txt/2202.10177.txt

CELLNUCLEI CLASSIFICATION INHISTOPATHOLOGICAL
IMAGES USING HYBRID OLCONV NET
Suvidha Tripathi
Department of Information Technology
Indian Institute of Information Technology Allahabad
Jhalwa, Deoghat, Prayagraj, Uttar Pradesh 211015
[email protected] Kumar Singh
Department of Information Technology
Indian Institute of Information Technology Allahabad
Jhalwa, Deoghat, Prayagraj, Uttar Pradesh 211015
[email protected]
February 22, 2022
ABSTRACT
Computer-aided histopathological image analysis for cancer detection is a major research challenge
in the medical domain. Automatic detection and classification of nuclei for cancer diagnosis impose
a lot of challenges in developing state of the art algorithms due to the heterogeneity of cell nuclei
and data set variability. Recently, a multitude of classification algorithms has used complex deep
learning models for their dataset. However, most of these methods are rigid and their architectural
arrangement suffers from inflexibility and non-interpretability. In this research article, we have pro-
posed a hybrid and flexible deep learning architecture O LConvNet that integrates the interpretability
of traditional object-level features and generalization of deep learning features by using a shallower
Convolutional Neural Network (CNN) named as CNN 3L.CNN 3Lreduces the training time by
training fewer parameters and hence eliminating space constraints imposed by deeper algorithms.
We used F1-score and multiclass Area Under the Curve (AUC) performance parameters to compare
the results. To further strengthen the viability of our architectural approach, we tested our proposed
methodology with state of the art deep learning architectures AlexNet, VGG16, VGG19, ResNet50,
InceptionV3, and DenseNet121 as backbone networks. After a comprehensive analysis of classi-
fication results from all four architectures, we observed that our proposed model works well and
perform better than contemporary complex algorithms.
Keywords Deep Learning, Hybrid networks, Object level features, Transfer Learning, Histopathological Images,
Cell Nuclei Classification, Class balancing, Convolutional Neural Networks, Multi Layer Perceptron
1 Introduction
Early cancer detection is a major challenge in the medical domain. Even today the medical community is largely
dependent upon the expert pathologist for detecting and classifying such cell anomalies that cause cancer, in whole
slide histopathological images. The job of the pathologists becomes very cumbersome and may take several days for
annotating the whole slide images of biopsy samples. Moreover, the reliability of predictions also depends upon the
experience of the pathologist and some times, consensus of more than one pathologists are required for confirming
such anomalies. These factors provide adequate motivation for research and development of a computer-assisted
diagnostic (CAD) systems which classifies cell nuclei and improves the understanding of some of the underlying
biological phenomenon, e.g., monitoring cancer cell cycle progress [1], type, shape, size and arrangement of the cells
in the affected organ sites, and the knowledge about metastasis, if the cells are present at some unlikely locations .
All these observations can be comprehended if we know the type of cell present in the diseased tissue sample. Early
diagnosis of cell anomalies can largely affect the disease prognosis [2]. Such as in the case of a colon or colorectal
carcinoma, epithelial cells lining the colon or rectum of the gastrointestinal tract are affected and timely detection of
these cells can help in quick diagnosis, which eventually would increase the prognostic value of the disease. Similarly,
the lymphocytes can also be analyzed for sentinel lymph node disease [2]. The other examples are the Myeloma orarXiv:2202.10177v1  [cs.CV]  21 Feb 2022APREPRINT - FEBRUARY 22, 2022
Figure 1: Example sub-images of different classes of nuclei starting from first row to fourth: Epithelial, Fibroblasts,
Inflammatory, Miscellaneous (in sets of 2 - Adipocyte, Endothelial, Mitotic Figure, and Necrotic Nucleus (from left
to right)
multiple Myeloma detections through the plasma cells which are the types of the white blood cells and cause the
cancer [3]. Therefore, a sample biopsy from a specific location can be quickly analyzed using the information of cell
environment provided by appropriate CAD system.
In particular medical image analysis, for all diagnosis, is attributed to the knowledge and skills possessed by trained and
experienced pathologists. Although pathologists have the ability and means to single out the affected cancerous lesions
in a tissue biopsy samples, most of such detections are still done manually and hence time-consuming. Numerous
challenges are involved in diagnosing cancer due to data set variability and heterogeneity in cell structures, which
makes the process extremely tedious even for experts. Software intervention for early detection is therefore important
for the purpose of effective control and treatment of the diseased organs [4]. To develop such automated cell detection
and classification algorithms, the knowledge of histology is vital and requires the annotated or labeled data set to
be prepared by the expert histo-pathologists. Once the labelled data is acquired, then the routine intervention of
pathologists can be eliminated while analyzing the whole slide samples under test by using the developed automated
CAD algorithms. Cell nuclei in a Hematoxylin and Eosin (H&E) stained histopathological slide sample have a specific
shade of blue caused by hematoxylin’s reaction with cellular protein present in the nuclei of the cells [5]. A shape
of the cell varies with cell type, cell-cycle stage, and also with the presence or absence of cancer. Fig. 1 shows four
different classes of nuclei namely, Inflammatory, Fibroblast, Epithelial, and Miscellaneous, which include adipocyte,
endothelial nucleus, mitotic figures, and necrotic nucleus [6]. The nuclei structures as shown in Fig.1 have different
shapes, texture, and intensity features, which vary by the factors i.e. nuclei type (epithelial, fibroblasts, lymphocytes,
etc.), the malignancy of the disease (or grade of cancer), and nuclei life cycle (interphase or mitotic phase) [7]. For
example, the inflammatory nuclei, a type of white blood cells and called as lymphocyte nuclei (LN) are smaller in size
and regular spherical shape as compared to epithelial nuclei (EN) [8]. Fibroblasts have long spindle-like shape and
appearance and having very little cytoplasm [9]. Activated fibroblasts and infiltrating inflammatory cells are the signs
of potential tumor growth [9]. All these histological and biological differences between cell structures and their site
of origin highlight the clinical relevance of classifying different types of nuclei.
In this paper, we have undertaken the feature based approach for automated nuclei classification. Feature based
approach can be classified into two general categories: hand-crafted features and deep learning based features. In
histopathology Images, morphological and architectural features whose accuracy depends on the amount of magnifi-
cation and the type of class and which has unique mixture of visual pattern qualify as hand crafted features whereas
unsupervised deep learning features are intuitive and are a by-product of filter responses obtained from large number
of training samples and fine tuning of the network. In the proposed work we clearly proved the benefits of using
combined feature set consisting both object level features and learned deep learning features over feature set acquired
from single domain on complex medical data. Moreover, for detailed analysis, the accuracy-generalization tradeoff
and space-time complexity issues as exhibited by traditional and DL methods respectively have been considered in the
proposed architectural arrangement.
In summary, key contributions of this work include:
1. The strength of our method lies in the flexible architecture that supports a backbone of deep learning model
to extract deep features and a simple object level extraction framework for extracting cell level features.
2APREPRINT - FEBRUARY 22, 2022
2. We achieved a high level of nuclei classification system through simple concatenation of derived features
from two domains.
3. The emphasis has been put through a series of experiments that in case of nuclei structures, even very small
number of basic and locally focussed object level features can enhance the performance if combined with
three or more layers of deep learning architecture.
4. To the best of our knowledge, this is the first study on this hypothesis that have developed a custom architec-
ture for the problem to highlight the need of designing lighter architectures for specific problems rather than
using deeper pre-trained architectures.
5. To the best of our knowledge, this is the first study on this hypothesis to have experimentally prove the
performance of end to end learning over stage wise learning.
The rest of the paper is organized as follows. Section-II describes the reviewed literature for handcrafted and deep
features. Section-III describes the complete methodology of the proposed work. The experimental setup is elaborated
in Section-IV including the database and workflow. Section-V contains various results and necessary discussion.
Discussion section also justifies the appropriateness of the proposed method while the flexibility and robustness are
the points of concerns. Section-VI concludes the work presented in the paper followed by the acknowledgments and
references.
2 Reviewed Literature
Owing to the above-mentioned properties exhibited by cell nuclei, many traditional handcrafted cell nuclei classi-
fication algorithms have been reported in [10–16]. Authors in [10] have first segmented the nuclei objects using
morphological region growing, wavelet decomposition and then found the shape and texture features for classifying
cancer cells vs. normal cells using SVM classifier. Another handcrafted feature-based method for cell nuclei clas-
sification in histopathological images while using shape, statistical, and texture (Gabor and Markov Random Field)
features from localized cells has been reported in [11]. Other object level (OL) feature extraction from localized cell
objects based methods have been reported in [12]. All the methods [10–12, 15, 16] using the OL features have been
critically analyzed against the utility of those features for individual problems related to a histological and/or cytologi-
cal analysis of cancer cells in [13,14]. The quality of the extracted features from various handcrafted methods [10–16]
is then assessed after passing them through appropriate classifiers. The success of their findings motivated the use
of targeted OL features in our methodology. To design effective handcrafted feature-based models requires complex
algorithms to achieve high performance and a decent level of domain-specific knowledge [15, 16]. Moreover, it be-
comes extremely difficult to resolve the issues due to dataset variability and heterogeneity within each cell type. These
issues lead to the inability of the reported novel but complex models to generalize well with varying datasets. It is
worth mentioning that, most of these methods are reported on a very small sample size in general, causing robustness
issues. To overcome the generalization problem, it is required to model features, which are common for a particular
class of cell nuclei but highly discriminating among different classes. Recently, deep learning architectures have been
known to produce generalized feature sets and hence have proved their niche in classification algorithms [17–24].
To put it more clearly, the key advantage of using deep learning architectures can be explained by highlighting the
problems of linear or other shallow classifiers. The traditional classifiers do not use raw pixel data and possibly cannot
distinguish two similar objects on different backgrounds which is a case of selectivity-invariance dilemma [17]. That
is why we need good feature extractors with such classifiers to solve the selectivity-invariance dilemma. The Deep
Learning (DL) based architectures automatically learn the good feature sets from the large histopathological image
data sets. In 2016, the CAMELYON challenge has also reported the use of extensive deep learning architectures for
solving various problems of Localization and Classification. Detailed methodologies of all these methods have been
reported in [21]. More recently, authors in [25] have used pre-trained VGG19 to classify extensively augmented multi
grade brain tumour samples whereas authors in [26] to identify alcoholism in subjects using their brain scans. So,
DL methods find applicability in wide range of applications due to their robust and better performing architectures.
However, there are some issues with deep learning based methods as well. DL features lack interpretability and can-
not be confirmed as global or local features. Moreover, there is always the lack of a large number of datasets in the
medical domain, which hampers or restricts DL algorithms to scale well on all other test data sets not used for the
training. Another major issue with deep architectures is the huge parameters with greater depths, causing the opti-
mization problem very time-consuming. At the same time, the complexity of a model increases, as the depth increases
and eventually the intermediate processes become less and less interpretable. One of the approaches for minimizing
the training time on the medical dataset is to use the concept of transfer learning and fine-tuning pre-trained models
such as AlexNet [18], VGG16, VGG19 [20], ResNet50 [19], DenseNet [27], and InceptionV3 [28]. Originally these
models have been trained on natural image datasets, which is from an entirely different domain but can be fine-tuned to
extract features on medical images. But, medical data has very little to no correspondence with natural images. Hence,
3APREPRINT - FEBRUARY 22, 2022
relying solely on transfer learning based fine-tune approach should not be preferred. Rather, the training should be
done on the networks that have either not been pre-trained on natural images or have been pre-trained on similar medi-
cal dataset. But, training DL networks on medical dataset has its own set of challenges, including lack of huge amount
of annotated medical data for training. Moreover, the diverse nature of medical images prevents the generalization and
standardization of datasets on which DL networks could be trained for transfer learning.
An exhaustive survey of deep learning methods as reported in [29] thoroughly highlights the merits of applying DL
methods in the field of medical imaging, medical informatics, translational bioinformatics, and public health. The
amalgamated use of both OL and DL features for the purpose of nuclei detection, segmentation, and classification
have also been suggested in [8, 29].
Therefore, OL features in combination with DL features could help to bridge the gap between issues that two domains
bring individually. Some recent articles have worked on the similar hypothesis of inter-domain feature combination
and developed a method that combines the two feature sets as reported in [22,23]. But, the drawback of their method is
in their complexity and huge training times due to very deep network models. Authors in [22] combined different deep
learning features extracted from Caffe-ref [30], VGG-f [31] and VGG19 models with Bag of Features (BoF) and Local
Binary Pattern (LBP) features. They then used ensemble classifiers to produce better classification accuracy than that
of the softmax classification method used by deep learning models. However, the dataset under experiments in [22]
was imbalanced hence, the reported accuracy trend may not hold good for other imbalanced datasets which are highly
probable in case of medical image datasets. F1 score and AUC are better parameters for assessing the performance of
classification algorithms for imbalanced datasets. Also, the authors [22, 23] reported the complex models which were
based on pre-trained deep architectures with 7 and more layers and did not analyze the performance trend on other
customized architectures that could have minimized the space and time constraints. It is difficult to design and test
such relatively inflexible algorithms on a new dataset and deploy in real time applications. For example, it is difficult
to change the design if one wishes to add a new functionality and re-train the algorithm. Furthermore, the reported
handcrafted features in these studies lack direct relevance to the nuclei structural properties.
3 METHODOLOGY
Hybrid feature based flexible classification framework trained on a dataset from [32] is used to determine the suit-
ability of combining different feature sets. Few pre-processing steps are performed to segment the cell nuclei from
background stroma. This step is necessary to extract the OL features. This feature set comprises relevant visual, shape
and texture features from each nucleus. DL feature is extracted from the original input images. Both the set of features
are then fused to produce a final feature set. The final fused feature set has been used by Multi-Layer Perceptron
(MLP) as input for classifying the cell nuclei into one of the four categories. The block diagram of the proposed
architectural setup has been shown in Fig. 2.
The entire flow has been modeled in the Algorithm-1.
Various steps involved in the proposed methodology has been elaborated in the following sub-sections (3.1)-(3.5)
3.1 Segmentation
Cytologic and histologic images prevent the generalization of segmentation algorithms because of the inherent vari-
ability of the nuclei structures present in them. Due to this reason, determining which state of the art algorithm for
nuclei segmentation would work for our dataset was a lengthy problem. Therefore, we seek to develop application-
specific segmentation algorithm for OL feature extraction. Our dataset contains H&E (Hematoxylin and Eosin) stained
RGB image blocks that stains the nuclei region in bright blue and cell region in pink. The staining helped us to roughly
extract the nucleus contour. Segmentation of an object then allowed for the calculation of OL features such as homo-
geneous color, texture, size, and shape of the segmented region.
Firstly, we enhanced the blue intensity of the nuclei through contrast adjustment. For this purpose blue color channel
intensities were mapped from initial values to 255. Similarly, Red and Green channel pixel values less than a certain
range were also tweaked towards higher range. This technique of adjusting intensity values in each channel to new
values in an output image helped in highlighting poorly contrasted nuclei regions from cell cytoplasm and background
noise. We assigned a higher value to blue intensity pixels relative to red and green components because blue-ratio is
proven to be capable of highlighting nuclei regions in H&E stained histopathological images [33]. This step is followed
by color normalization so that the intensity values follow normal distribution and as well remove any noise/artefact
that may have been introduced due to contrast enhancement. For the next step, we computed the binary image and
calculated the convex hull of the labelled region having the highest number of pixels. Convex hull of the binary
image ensured that the largest area containing most blue pixels is retained and the defined boundary of the nuclei can
4APREPRINT - FEBRUARY 22, 2022
Figure 2: Block Diagram of proposed OLConvNet. Raw training images of cell nuclei is passed through Branch-1 of
the network for DL feature extraction and further for classification using fully connected (FC1) and softmax layer of
the DL network (OUTPUT-1). OL features are extracted from segmented nuclei images after segmentation pipeline.
Extracted OL features are classified in Branch 2. Switch between branch 1 and 2 make the decision about which kind
of output we would want for our dataset (OUTPUT-1 or OUTPUT-2 or Both).
be obtained for calculating OL features. In other words, the perturbation in nuclei structure due to staining process
may distort original nuclei structures, so obtaining a convex hull defines the smooth boundary around the nucleus.
This further helps in following procedural steps of extracting OL features. Convex hull step is then followed by edge
extraction of the convex hull. Lastly, we did the scalar multiplication of the resultant image with the original image to
obtain the final output of a segmented RGB Nuclear image. Segmentation results helped in delineating nuclei region
from the surrounding tissues. Figure 3 shows the pipeline of segmentation. Some of the segmented classwise nuclei
examples are shown in Figure 4.
Figure 3: Segmentation Pipeline of our network
5APREPRINT - FEBRUARY 22, 2022
Algorithm 1 OLConvNet
Input: Training data set Dtrwith msamples. Data fi(X), where i= 1;_:::; m is an instance in the 3 dimensional
image space X2<UVdandyi2Y=f1;2;3;4gis the class identity label associated with fi(X).
Output: Output1: ^ycnn1-of-4 Softmax output from CNN 3L, Output2: ^ymlp1-of-4 Softmax output from MLP
Pre - Processing (Balancing Dataset) :ffi(X); yigM ADASY N (ffi(X); yigm).M - samples after balancing
dataset, ADASYN - a function used for balancing the dataset
forxi2Xdo
SEGMENTATION :xs
i SegN (xi) .SegN: a function following segmentation pipeline
OBJECT LEVEL FEATURES :(FVi
OL)MN1 OLFV (xs
i)
.OLFV: a function extracting nuclei features, N1- Number of Object Level features
STAGEWISE - TRAINING :
forepoche2(1;100) do
forbatch b do
forxi2bdo
CNN Inference: ^yicnn CNN !cnn(xi)
@L
@!cnn ^ycnn
i;yi .!- weight parameter
Compute@L
@!cnnusing backpropagation .L - cross-entropy Loss function
Update CNN: !cnn !cnn@L
@!cnn .- Learning rate
ExtractFeatures:fFVcnng(MN2) CNN 3L(ff(X)g;layer ) .N 2- Number of CNN features
Concatenation:fTotalFeaturesgM(N1+N2) FVcnn+FVOL
forTFi2TotalFeatures do
MLP Inference: ^yimlp MLP!mlp(TFi)
@L
@!mlp ^ymlp
i;yi
Compute@L
@!mlpusing backpropagation
Update MLP: !mlp !mlp@L
@!mlp
END to END - TRAINING :
forepoche2(1;100) do
forbatch b do
forxi2bdo
CNN Inference: ^yicnn CNN !cnn(xi)
@L
@!cnn ^ycnn
i;yi
Compute@L
@!cnnusing backpropagation
Update CNN: !cnn !cnn@L
@!cnn
ExtractFeat:fFVi
cnng(1N2) CNN 3L(ffi(X)g;layer )
Concatenation:fTotalFeaturesig1(N1+N2) FVi
cnn+FVi
OL
MLP Inference: ^yimlp MLP!mlp(TotalFeaturesi)
@L
@!mlp ^ymlp
i;yi
Compute@L
@!mlpusing backpropagation
Update MLP: !mlp !mlp@L
@!mlp
Figure 4: Example segmented-images of different classes of nuclei starting from first row to fourth: Epithelial nuclei,
Fibroblasts, Inflammatory nuclei, and Miscellaneous
6APREPRINT - FEBRUARY 22, 2022
3.2 Object Level Feature Extraction
We have extracted a set of nine features. These nine features include color, texture and shape information of the nuclei
calculated on the segmented nuclei image. Color information contains the pixel intensity value that has the highest
frequency in the histogram. Since the nuclei of the cells take the shades of blue after H&E staining, the area which
has the maximum intensity of blue would be the area inside the nucleus. If the intensity with maximum frequency
is not around the blue range, it is not the nucleus and that acts as the differencing factor while classifying nuclei
against miscellaneous or poorly segmented region. Texture information contains the GLCM texture features [34].
Among many GLCM texture features, we calculated four statistical texture features which are Contrast, Homogeneity,
Correlation, and Energy of the nucleus surface. Texture, being a fundamental property of tissue surfaces, helps to
differentiate between different types of cells such as epithelial, fibroblasts and lymph node cells. In papers [9], [11]
and [12] authors described the shape and morphology features of different classes of nuclei. The considered shape and
morphological features in this paper are the areas, major axis length, minor axis length, and eccentricity.
3.3 Convolutional Feature Extraction
CNN Architecture: The general CNN architecture first proposed by Yann LeCun et al. in their paper [35]. In CNN,
unlike traditional machine learning approaches that take features as input to classify the data into categories, raw
images are used as input. It works on the input images by learning filter weights and modify itself until convergence.
The basic architecture of CNN comprises seven major layers, namely, Input image layer, Convolution layer, ReLU
layer (Non-Linearity), Pooling (Local Max), Fully connected layer, Softmax Layer, and Classification Layer.
An exhaustive theoretical account of CNN can be found in [36]. These layers, when combined in a pattern, create
deeper networks with the increased number of hidden units. Deeper the network, a more exhaustive set of features
could be extracted from the image [37]. However, this theory is highly subjective and depends majorly on the properties
of the datasets. For example, the size of the image data should be large enough to be processed into a meaningful
representation in case of architectures with great depths. Also, in a few cases, the number of parameters resulting
due to large depths pose a major disadvantage in terms of computational load and efficiency of the algorithm. Hence,
we did experiments to develop a custom three layer convolutional network to extract features from the nuclei dataset.
We named it CNN 3Lfor ease of reference. Figure 2 shows the network in detail. After training the whole network
for a set number of epochs, the final network yields the best set of features which are then further progressed to the
classification layer. For the purpose of extracting DL features, it is preferred to extract features from the last layer
located before the fully connected layer since the final convolution layer features are a more specific representation
of the dataset images. Our proposed setup also has the flexibility to change the backbone architecture from shallow
three-layer network to any number of layers.
3.4 Fusion
We get two set of features which are globally extracted from segmented nuclei (OL features) and automatically ex-
tracted features that include both local and global information (DL features) of the image. We have concatenated
these two sets along the axis that contained the feature vector of a sample and produced one combined set for further
classification (Fig. 5). This exhaustive set of features are then used as input to the first MLP layer for the purpose of
categorizing them into four classes.
Figure 5: Diagram displaying fusion of OL and DL features
7APREPRINT - FEBRUARY 22, 2022
3.5 Classification
For classification, we have used Multi-Layer Perceptron (MLP). This process is called Transfer Learning where CNN
is used only for feature extraction while the next step of classification is performed by another machine learning
algorithm such as MLP for multiclass classification. In this paper, we used the MLP network with one input layer, one
hidden layer having ten nodes, and one output layer. Combination of features is fed as an input to the first MLP layer.
Hidden layer used tansig as an activation function defined by
tansig (n) =2
(1 +exp(2n))1; (1)
where nis our input vector to the second hidden layer. This activation function is faster than tanh , however numerical
differences are small [38]. The output prediction scores from the output layer of MLP was given by softmax function
[39] - [40] defined as:
pj(^y(x)) =exp(^yj(x))P
kexp(^yk(x)); (2)
where ^yj(x)denotes the jthelement of output vector ^y(x)We performed 2-fold cross-validation test on our network
to observe the efficacy of our model.
4 Experimental Setup
4.1 Database
For our experiments, we have taken the database from [32]. This database has a total of 100 H&E stained histology
images of size 500 by 500 pixels. The center coordinate location of each nucleus in each image has been provided
as ground truth. Different types of nuclei are divided into sets belonging to four classes which are Epithelial nuclei,
Fibroblast nuclei, Inflammatory nuclei and the rest of the small types are categorized as one class called ’miscella-
neous’. We obtain subimages of size 27x27 extracted around the center coordinate locations and maintained them
in four sub-folders segregated by class types. Four classes in the dataset have 7,722 Epithelial, 5,712 Fibroblasts,
6,970 Inflammatory, and 2039 miscellaneous nuclei totaling up to 22,444 nuclei in 100 H&E stained histopathological
images. The number of samples in each class are imbalanced which can cause classification results biased towards the
majority class ( the class having the highest number of samples). A systematic study of the class imbalance problem
in convolutional neural networks by authors in [41] have concluded through their research that the class imbalance
problem, if not addressed may have a detrimental effect on classification performance. The influence of imbalance
on classification performance increases with the size of the dataset. They also reported that in case of convolutional
neural networks, the method that works best for eliminating class imbalance problem is oversampling with thresh-
olding which opposed to other machine learning models, does not cause overfitting in CNN. Thus, we performed
our experiments after balancing our dataset. We used adaptive synthetic sampling for eliminating class imbalance by
synthetically creating new examples in the vicinity of the boundary between the two classes than in the interior of
the minority class via linear interpolation between existing minority class examples [42]. So, after creating synthetic
examples for minority classes (Fibroblast, Inflammatory, and miscellaneous) with respect to number of samples in
majority class (Epithelial) we accumulated 29,771 total data points having 7,722, 7,275, 6,970, and 7804 samples
in class 1 (Epithelial), class 2 (Fibroblast), class 3 (Inflammatory), and class 4 (Miscellaneous), respectively. After
acquiring the balanced nuclei dataset, for the purpose of training, validation, and testing, we divided each class in
the ratio 0.7:0.15:0.15, respectively. All the networks are evaluated on the testing set and performance metrics are
reflected in Section V .
4.2 CNN 3Land Parameter Settings
We used the CNN framework with three convolutional layers and one fully connected layer as shown in Figure 2.
Traditionally, CONV RELU layers are stacked, which are then followed by POOL layer. This pattern is repeated until
we get a relatively small number of parameters from the image or when the optimal performance is achieved. The
aim to stack up layers is to extract a more exhaustive set of features. The last fully-connected layer holds the output,
such as the class scores. In our research work, experiments with a different number of layers yielded the optimal value
of three convolution layers and one fully connected layer. We have used our knowledge about the problem, applied
a heuristic approach to select parameters, and observed the outputs. Based on the outputs obtained, we tweaked
the parameters. This process was repeated until we got the optimal set. The architecture was trained with different
hyperparameters such as the number of convolutional layers and fully connected layers, learning rate and the number
of epochs. The accuracy achieved at each instance was recorded and plotted as shown in figures 6a and 6b. The plots
8APREPRINT - FEBRUARY 22, 2022
show the accuracy values on the Y-axis corresponding to the number of epochs on X-axis for each type of architecture
taken into consideration. Here, NCMFCdenotes N Convolutional layers and M Fully Connected layers. The two
figures show observations recorded from two learning rates.
(a) Accuracy at learning rate 1e-4
 (b) Accuracy at learning rate 1e-5
Figure 6: Epochs v/s Accuracy graph at different hyperparameters
From the experiments, we found out that neither stacking up more layers nor decreasing the learning rate showed a
positive impact on the performance of the network. A learning rate of 1e4with three convolutional layers and 1
fully connected layer the accuracy recorded at epoch number 100 was the highest (refer figure 6a). Increasing the
learning rate further to 1e3for CNN 3Ldecreased the accuracy to 61.77% at epoch 70 and further down to 59.41%
at epoch 100.
For training the final optimal network (CNN 3L), we normalized our dataset images by subtracting the mean image of
the training set from all training set images before using them as input vectors. If we don’t scale our input training
vectors, the feature values distribution range would likely be different for each feature, and thus the corrections in each
dimension after each epoch through learning rate will differ for each dimension. This might cause overcompensation
(very high variance from mean weight) in weight of one dimension while undercompensating (very low variance from
mean weight) the another. This creates a non-ideal situation since it might get difficult to center the weight space. This
will also affect the time efficiency as it will travel too slow to get to a better maxima.
In our architecture CNN 3L, We followed 5 x 5, 100 channel, 3 x 3, 50 channels, and 3 x 3, 100 channels for each
convolutional layer respectively. We experimented with different filter window size and numbers of filters and ob-
served little to no improvement in the performance, for example, experiment with filter numbers 32, 64 and 128 on
subsequent three convolutional layers of the optimized CNN 3Lachieved the F1-score of 0.7907. There could be an
infinite number of permutations to choose network parameters from. We performed numerous experiments to decide
on CNN 3Las our optimal network, however finding a theoretical justification is out of the scope of this work. Convo-
lution layer is followed by ReLU layer. There is no change in the dimension of the output of this layer. Next is the Max
pooling layer with size 2 and stride 1. Since, our network depth is less so, keeping stride 1 in Max pooling layer was a
design choice to avoid loss of information. Total number of learnable parameters were only 118,000 as compared to
heavier networks like AlexNet ( 56,000,000), VGG16 ( 134,000,000), VGG19( 139,000,000), ResNet50( 23,000,000),
InceptionV3( 22,000,000), and DenseNet121( 7,000,000). After fixing the network architecture, first branch 1 of the
network was trained. Trained Branch 1 (CNN layers) were tested on the test set and the results ( OUTPUT -1 ) are
accumulated in Table 1. Applying transfer learning approach [43], we used MLP for final classification outcome
(OUTPUT - 2) instead of 1-of-4 classification output from a softmax layer. Trained CNN backbone has been used
as fixed feature extractor. The approach is stage-wise learning and the output 2 observations are recorded in Table 2
under the column named as ”stage-wise”.
4.3 End-to-End Training
We performed end-to-end training of our three layer CNN architecture (CNN 3Lto show the performance of our net-
work when the learning happened from end-to-end rather than stage wise. By end-to-end learning, we mean that both
CNN network (CNN 3L) and MLP network were trained together using the intermediate outputs of CNN 3Las inputs to
the MLP network. So, MLP network received two sets of input features, activated features from the last Max pooling
layer of CNN 3Land the other from the pre-calculated handcrafted features. These two set of features were concate-
nated before giving them as input to the MLP layers. At the end of the training, two outputs were accumulated, one
9APREPRINT - FEBRUARY 22, 2022
from the CNN 3L1 of 4 softmax outputs and another from the MLP layer classification. The diagram showing the
architectural setup is shown in Figure 2. We performed two experiments in different settings.
1. Branch 1 (CNN layers) of the network (refer Figure 2) was trained keeping branch 2 (MLP layers) inactive
by turning off the switch and the results of 1-of-4 classification layer of CNN was recorded.
2. Both branch 1 and branch 2 were trained simultaneously to obtain two outputs, one from CNN layers
(OUTPUT-1) and second from MLP layers (OUTPUT-2) (refer Figure 2). The switch was connected in
this case.
We have reported the output of MLP classification of combined feature set (OUTPUT-2) in Table 2.
5 Results and Discussion
The aim of this work is to highlight the different outcomes of classifying the database with different settings. To
represent the outcomes in order of increasing classification performance we have divided our results in the following
categories.
1. Classification output of only deep learning features from softmax layer of CNN 3L
2. Classification results of combined feature set using MLP (O LConvNet) - Stagewise.
3. Classification results of combined feature set using MLP (O LConvNet) - End-to-end.
5.1 Results
Data imbalance has been eliminated using the method described in Section 4.1. We evaluated the proposed model
on the basis of F1-scores and multiclass AUC. Multiclass AUC is calculated using the prediction scores given by the
softmax function described in (2).
Table 1 shows the comparative classification performance of MLP and softmax CNN networks, separately. OL features
classified the nuclei classes using MLP classifier (Branch - 2 of Fig. 2) while DL features were propagated along the
fully connected and classification layers of CNN networks (Branch - 1 of Fig. 2). To visualize the network flow
in both OL classification and deep learning classification, connection through the switch in Fig. 2 was broken to
create two separate branches where, both the branches were trained mutually exclusively. We compared the OL
features performance with CRImage [44] (Table 1) which also calculates features after the segmentation of the nuclei
using basic thresholding, morphological operations, distance transform, and watershed. CRImage also calculates
statistical features from the segmented nucleus but lacks visual, shape and texture features. Besides statistical features,
incorporation of histopathologically deduced features such as nuclei size, shape, area, color, and texture hold direct
relevance to the dataset and therefore the absence of such features prevents CRImage to perform well on the dataset.
So, from the observations reported in Table 1, we deduced that our OL features, though only a small number, performed
better than [44]. Deep models CNN 3L, AlexNet, VGG16, VGG19, ResNet50, InceptionV3, and DenseNet121 when
tested (without OL features) on the dataset produced F1-score and AUC better than OL features by a huge margin due
to exhaustive feature set produced after convolution operations. These features are non-interpretable but have been
known to perform quite well in classification tasks. Hence, Table 1 reports the individual performance of OL features
and DL features and proves the statements made in the ’Introduction’ Section about the need to shift from traditional
feature engineering to convolutional feature learning.
Next, we trained the network in Fig. 2 stage wise. Features from trained deep learning networks from the previous
experiment were concatenated with OL handcrafted features. Then, the second stage of the training of concatenated
features was done by MLP classifier. The results obtained on test dataset after MLP training was then reported as
combined features classification performance. End to end training followed stagewise experiments to analyze the
effect of the performance of our network when trained in two ways. We have shown the difference through F1-score,
multiclass AUC and cross-entropy loss in Table 2 which reports performance of classification on a 2-fold cross-
validation experiment. The obtained results from stage-wise training, as expected, showed improvement in F1-Score
and Multiclass AUC in comparison to individual results obtained after classifying with only DL features and only OL
features (Table 1). While there is only a 2% increase in F1-score and 1% increase in AUC score in case of CNN 3L,
a marked improvement has been recorded in case of deeper pre-trained architectures. Whereas, in the case of end-2-
end training, no improvement in the performance metrics has been observed in any of the backbone networks. Also,
the higher loss value recorded in the case of end-2-end approach reflects decreased performance in comparison to
the stage-wise approach. These cross-entropy loss values highlight that the joint loss propagation after the shared
layer (concatenation layer) might have affected the overall performance of the model. The additional OL feature
10APREPRINT - FEBRUARY 22, 2022
Table 1: Comparison between methods stratified by classifier
MethodPrecision Recall F1-Score Multiclass AUCBackbone Classifier
Only Object Level MLP 0.5154 0.5156 0.5135 0.7857
CRImage [44] SVM (RBF kernel) * * 0.4880 0.6840
CNN 3L Softmax 0.8043 0.8046 0.8040 0.9441
AlexNet Softmax 0.8280 0.8281 0.8216 0.9386
VGG16 Softmax 0.8693 0.8699 0.8689 0.9757
VGG19 Softmax 0.8575 0.8581 0.8578 0.9701
ResNet50 Softmax 0.8900 0.8893 0.8892 0.9799
InceptionV3 Softmax 0.8164 0.8175 0.8175 0.9538
DenseNet121 Softmax 0.8784 0.8706 0.8706 0.9756* data
unavailable
set concatenated simultaneously during training on the shared concatenation layer did not improve the discriminative
property of the DL features and hence, the subsequent MLP layers performed in the same way as the fully connected
and softmax layers of DL models.
Fig. 7 show the ROC curves obtained after end-2-end training on all four networks. The subfigures 7a, 7b, 7c, 7d,
7e, 7f, and 7g shows the ROC curves of all four classes with respect to seven backbone networks, discriminated
through four colors. The figure also mentions the AUC value for individual classes. Micro and Macro average of
four ROC curves from each class show similar value because our dataset is balanced. The micro-average method
calculates the sum of true positives, false positives, and false negatives for different sets whereas, in Macro-average
method, the average of the precision and recall is taken into account. Micro Average is preferred when there is a
class imbalance problem. The Figure also labels AUC values for each class. After recording the values we observed
a dip of 2% to 4% for class 2 (Fibroblast) and 3 (Inflammatory), whereas, the AUC values for classes 1 (Epithelial)
and 4 (Miscellaneous) are comparable across all four networks. The decrease in performance can be attributed to
the number of samples which are relatively less for class 2 (7275) and class 3 (6970) than class 1 (7722) and class
4 (7804). Also, in the case of Fibroblasts (class 2), the long spindle-shaped cell barely has a visible nucleus which
prevents segmentation algorithms to detect the nucleus area effectively. Consistent class-wise performance across all
four backbones and high AUC are some of the other observations deduced from Figure 7.
2-fold cross validation performance metrics with CNN 3Lalong with deep architectures AlexNet H, VGG16 H, VGG19 H,
ResNet50 H, InceptionV3 H,and DenseNet121 H, where H stands for Hybrid, are reported in Table 3 and Figure 8. The
OLConvNet metrics recorded in Table 3 are stagewise observations from OUTPUT - 2 of the network obtained after
combined feature set testing. We compared our results with Softmax CNN + SSPP (Standard Single Patch Predictor)
and Softmax CNN + NEP (Neighboring Ensemble Predictor) [32] architectures used for classification of nuclei from
the nuclei dataset cited by [32]. The authors in this article worked on the theory that the pixel of interest is likely
the center of the nucleus and using this theory they formulated the classification algorithm by spatially constraining
the high probability pixels in the vicinity of the centers of nuclei. They proposed two algorithms called Neighboring
Ensemble Predictor (NEP) and Standard Single Patch Predictor(SSPP) which when coupled with SC-CNN (Spatially
Constrained- CNN) produced the classification results as mentioned in Table 2. The drawback of their method was
in building a complex target specific model which lead to low classification performance. Their CNN architecture is
custom but they didn’t test their methodology on deeper pre-trained architectures which might have performed better
with or without their elaborate model. Further comparison with superpixel descriptor [15] and CRImage on the same
dataset show that both the methods performed relatively poor and that our method exhibited a higher performance
regardless of the backbone architecture used. The motivation behind devising Superpixel descriptor was to distinguish
the area with the different histologic pattern. This descriptor lacks direct features related to the visual appearance of
the nucleus, like color, texture, and shape, thus yielded lower classification performance. Whereas, CRImage [24] only
calculates segmented nuclei features. Segmented nuclei features are insufficient to classify complex nuclei patterns
due to either weak staining or presence of overlapping boundaries. This prevents CRImage to perform well in this
dataset. We have shown the same through our only OL features classification results in Table 1. Figure 8 reports
comparative class-wise classification performance of all the methods. F1-score of Miscellaneous class show a marked
jump in values obtained from O LConvNet and an improvement of more than 47% is recorded between the highest
and lowest performing algorithms which are ResNet50 and CRImage, respectively. For Epithelial and Fibroblast, the
11APREPRINT - FEBRUARY 22, 2022
(a) CNN 3LTest ROC
 (b) AlexNet Test ROC
(c) VGG16 Test ROC
 (d) VGG19 Test ROC
(e) ResNet50 Test ROC
 (f) InceptionV3 Test ROC
(g) DenseNet121 Test ROC
Figure 7: ROC curves from four backbone networks using O LConvNet (end-2-end learning)
12APREPRINT - FEBRUARY 22, 2022
Table 2: Performance parameters from stage wise and end-2-end learning experiments (OUTPUT - 2)
Method BackboneStagewise End-2-End
F1-score AUC loss F1-score AUC loss
OLConvNetCNN 3L 0.8243 0.9587 0.1211 0.8084 0.9500 1.2750
AlexNet H 0.9542 0.9903 0.0915 0.8359 0.9600 1.0860
VGG16 H 0.9569 0.9923 0.1036 0.8731 0.9700 0.9407
VGG19 H 0.9546 0.9879 0.1221 0.8675 0.9700 1.1213
ResNet50 H 0.9677 0.9973 0.0272 0.8892 0.9799 0.7335
InceptionV3 H 0.9618 0.9963 0.0309 0.8175 0.9538 1.0795
DenseNet121 H 0.9616 0.9961 0.0318 0.8706 0.9756 0.7972
class wise performance of CNN 3Lwas marginally lesser than SSPP but, a difference of 4% to 7% was recorded
with NEP whereas, in case of DL models used in O LConvNet, the F1-scores have been consistently better than all
the algorithms used for comparison. Same can be interpreted from Figure 8 where the classwise performance of
OLConvNet backbones is consistently better than contemporary algorithms.
Table 3: Comparative Performance parameters for Nucleus Classification
Method Backbone Precision Recall F1-Score Multiclass AUC
Softmax CNN + SSPP [32] CNN * * 0.7480 0.8930
Softmax CNN + NEP [32] CNN * * 0.7840 0.9170
Superpixel Descriptor [15] - * * 0.6870 0.8530
CRImage [44] - * * 0.4880 0.6840
OLConvNetCNN 3L 0.8241 0.8245 0.8243 0.9587
AlexNet H 0.9578 0.9577 0.9578 0.9953
VGG16 H 0.9611 0.9610 0.9610 0.9960
VGG19 H 0.9544 0.9548 0.9546 0.9879
ResNet50 H 0.9676 0.9678 0.9677 0.9973
InceptionV3 H 0.9618 0.9618 0.9618 0.9963
DenseNet H 0.9616 0.9617 0.9616 0.9961* data
unavailable
5.2 Observations and Discussions
The experiments performed in this study gave some interesting points for discussion. Formally, whenever we per-
formed deep learning based classification tasks, our focus generally remained on improving the classification perfor-
mance of the architecture by fine-tuning or transfer learning. Transfer learning is generally used in the case when there
are fewer samples of similar data. In our case as well, samples were not enough for deep architectures to generalize
well on the dataset. Moreover, pre-trained weights of discriminative nuclei types to transfer learn on our database was
also not available. Also, the intention to not use pre-trained publically available ImageNet weights was because our
dataset is very dissimilar to ImageNet. Hence, we chose to build our own CNN layers (CNN 3L) and fine-tune them to
produce optimized performance parameters.
DL features from CNN 3Land OL features representing visual and structural properties of nuclei were concatenated
to produce a combined feature set which consequently improved classification results. F1-Score and multiclass AUC
values recorded in Table 1 reflected the individual performance of the traditional handcrafted OL features with MLP
classifier and DL features with Softmax classifier whereas, F1-Score and multiclass AUC values in Table 3 show
the performance of the O LConvNet network with combined feature sets. These observations reflect that even if the
OL feature set was very small (total 9) relative to CNN 3Lfeature length (48,400) , a marked difference between 1-
of-4 softmax classification (OUTPUT 1 of Fig. 2) and O LConvNet classification (OUTPUT 2 of Fig. 2) reflect the
applicability and importance of including object-specific features. Further, the high performance metric values in Table
3, exhibited by deeper pre-trained architectures ( AlexNet H,V GG 16H,V GG 19H,ResNet 50H,InceptionV 3H,
13APREPRINT - FEBRUARY 22, 2022
Figure 8: Comparative results for nucleus classification stratified with respect to class label
DenseNet 121H) shows that increasing depth of the network although increase the capability to extract discriminative
DL features and hence better F1-score and AUC from CNN 3L(as reflected in Table 7) but, the discriminative ability of
the DL feature set is enhanced many folds when combined with OL features. This statement can be verified from Table
2 which shows the increase of 10% in F1-score and AUC values of these pre-trained architectures. However, training
OLConvNet jointly for classification (end-2-end ) ( Table 2) using a combined feature set did not affect F1-score
and AUC values. This is likely because we trained the network with the combined loss of CNN classification layer
and MLP output layer. The combined cross-entropy loss affected the performance of the joint network. Therefore,
from Table 2 we observe that the loss values are higher in end-2-end training in comparison to stage-wise training.
Hence, stagewise 2-fold cross-validation results are preferred over end-2-end classification. The results were also
suggestive of the fact that the classification performance has most certainly increased quite dramatically as the levels
of convolution layers increased from just three layers in CNN 3Lto 121 layers in DenseNet121. In all the cases, though,
our hypothesis remained true and our architectural setup provided the room to make modifications in the configuration
of backbone networks.
The question that can be raised at this point is why to use CNN 3Lwhen we are getting better performance with deeper
architectures. In support of CNN 3L, when we look in Table 2 and observe cross-entropy loss of all backbones in
stagewise learning, we see that CNN 3Lhas comparable loss value with deeper networks but, the number of parameters
and the training time taken by AlexNet, VGG16, VGG19, ResNet50, InceptionV3, and DenseNet121 shown in Table
4 is much larger than CNN 3L. This opens up the discussion for the need of using an optimal number of layers to
achieve satisfactory performance by an application instead of very deep CNN models. To further strengthen our point,
we know that in deep learning, a number of successful early and recent state of the art architectures such as AlexNet,
VGG16, VGG19, , ResNet50, InceptionV3, and DenseNet121 etc., were developed to increase the abstractness in
extracting deep and better quality feature maps, primarily for the purpose of classification. However, the standard
datasets used to test these networks were real life natural images.
Medical data is a completely different modality that has high variations in features from the same class. Therefore,
the state of the art deep network models on such datasets do not perform well and suffer from high validation and
testing loss. Another main reason for their failure is a scarcity of labeled data volume in the medical image domain.
To cope up with these limitations, every new model that is being developed for medical image data classification
is function specific and does not address the global problems in the domain. Most of the recent literature in the
classification of structures in histology images do not use raw images for deep learning models. A fair amount of pre
and post processing of data is required to enhance the performance which consequently, hampers the generalization
capability of the model. So, instead of building novel models with less or limited applicability in the cell nuclei image
classification problem, it seemed better to change the way these models are used. With simple architectural change
and introduction of visually and structurally relevant handcrafted feature set, through experiments, we have established
gain in performance values. Our model is flexible in a way that any deep model can be fitted in the architecture to
extract deep features. Another highlight of our work is to show how the addition of a handful of basic OL handcrafted
14APREPRINT - FEBRUARY 22, 2022
features can bring a notable change in the final classification output. We do not need to design special descriptors or
use complex handcrafted features to complement deep learning features. Instead, a small number of discriminative
OL features in case of nuclei classification like color, texture, and shape can enhance the discriminative capability of
the neural network classifier.
While we agree that the architecture is simple, we first need a fine-tuned and flexible model that scales well with
the number of options of novel models available today. More than that, we need a compact model with less training
time ( a comparatively shallow convolutional network ) that fairly works well and give comparable results with deeper
architectures. We compared the time taken by each of the methods to train in Table 4. We observed that CNN 3L
took the least amount of time and very huge differences were noted when compared with ResNet50, InceptionV3, and
DenseNet121. Time may be regarded as an insignificant paradigm with advanced computer systems available today.
However here, it is important to mention that, we trained our algorithms using three NVIDIA GeForce GTX 1080
Ti GPUs with 11GB RAM each, in parallel. This much computing power is not still really common in most of the
places and, also it becomes difficult to train heavy deep networks in light mobile applications or hand-held systems.
These much computing resources are impossible to accommodate in lighter applications as of yet and hence, shallow
networks that work fairly well for general diagnostic procedures can help in reducing space and time constraints for
such systems. In this research, we are dealing with time sensitive application where the quick delivery of classification
results are important to help pathologists to proceed for further analysis in cancerous tissues and diagnose cancer as
soon as possible. Besides time, we would also like to highlight that our architecture CNN 3Lis not using pre-trained
weights of ImageNet which has several other advantages such as, applications can use the custom size of their dataset
directly without the need of resizing it to conform according to the size of ImageNet data. This is beneficial at times
where very large or very small images may lose quite a significant amount of details when scaled. While one may
argue in this case that we could have trained all DL backbones from scratch without the need of using pre-trained
weights. However, these deep networks require a certain minimum size of the image (48 x 48 in case of VGG16 and
VGG19, 197 for resNet50 and 221 for DenseNet121) at the input layer to train. Hence, the limitation of these state
of the art deep learning architectures with complex datasets is too much to ignore. Additionaly, for training such deep
networks from scratch, we would need a very large number of dataset samples for the networks to learn efficiently.
Table 4 summarizes the time is taken and the number of trainable parameters used by CNN 3L, AlexNet, VGG16,
VGG19, ResNet50, InceptionV3, and DenseNet121 to train the dataset.
Table 4: Time and trainable parameters comparison between backbone architectures
Method No. of parameters Time (seconds)
CNN 3L 118 thousands 303
AlexNet 56 millions 3304
VGG16 134 millions 1408
VGG19 139 millions 1420
ResNet50 23 millions 18951
InceptionV3 22 millions 13314
DenseNet121 7 millions 24515
Hence, a simple architectural setup whose components can be modelled as per the dataset requirement and which uses
the combination of OL features and shallow, yet incorporating the properties of DL model - CNN 3Lin case of complex
dataset such as ours has proved to be a better approach than the traditional OL models or trending DL algorithms, alone
that do not allow changes in the architecture and require specific configurations to work well. It is important to note
that there are no standard datasets yet in case of histological nuclei images so, comparing various methods mentioned
in the literature on only one database does not guarantee to give expected results. According to the free lunch theorem,
which is most certainly applicable in case of medical image databases, there is no global algorithm that could be
developed to give good results across all kinds of histopathological data. Each experiment is conducted with different
dataset acquired by the research team themselves or from the pathologists, whose property changes with the location
of the disease. Their results are validated using different performance metrics. Hence, standard datasets and ground
truth in the case of complex histopathological cancer images is a current challenge in this field of research.
6 Conclusion
The knowledge about the cell of origin of a tumor may benefit doctors to treat the tumor more effectively since a correct
classification will greatly increase the biological understanding of carcinogenesis. Using this motivation in this paper,
15APREPRINT - FEBRUARY 22, 2022
we have built our classification model by emphasizing on a hybrid and flexible model that can incorporate the two
wide domains of features, the age-old traditional object-level features such as intensity, texture, and shape features and
recent deep learning based features. While object-level features have proved their efficiency in various domains and
purposes of biomedical image processing, including cancer-based disease recognition and classification, the current
trend has drastically shifted towards using various deep learning methods. Our work tried to highlight through the
results that using only deep learning might not work in case of all datasets. Therefore, a need to develop a shallow
yet effective architecture such as CNN 3Lincorporated in our proposed skeleton called O LConvNet, that could robustly
combine the benefits of object level features in this study, motivated our work. Moreover, to guarantee a wide range of
applicability, a model that is easy to understand, and deploy is required, also, which has the flexibility to incorporate
both deeper, more efficient networks like AlexNet, VGG16, VGG19, ResNet50, InceptionV3, and DenseNet121, and
shallow, light model like CNN 3Lso that the O LConvNet can be adapted to a wider range of applications. The results
were encouraging and our network performed better than the recent state of the art implementation on the same dataset.
The future work could be to incorporate better algorithms that combine the best of both the worlds i.e. traditional object
level and deep learning features. Approaches can differ with better classifiers as well. In conclusion, our method opens
the possibility of further research in developing more robust nuclei classification models that can scale well on all kind
of datasets.
7 Acknowledgments
This research was carried out in Indian Institute of Information Technology, Allahabad and supported by the Ministry
of Human Resource and Development, Government of India. We are also grateful to the NVIDIA corporation for
supporting our research in this area by granting us TitanX (PASCAL) GPU.
References
[1] Xiaowei Chen, Xiaobo Zhou, and Stephen TC Wong. Automated segmentation, classification, and tracking of
cancer cell nuclei in time-lapse microscopy. IEEE Transactions on Biomedical Engineering , 53(4):762–766,
2006.
[2] C ˆatoi C. Baba AI. Comparative Oncology , chapter 3. The Publishing House of the Romanian Academy,
Bucharest, 2007.
[3] Cancer Research UK. Types of cancer, November 2017.
[4] Anant Madabhushi and George Lee. Image analysis and machine learning in digital pathology: challenges and
opportunities, 2016.
[5] Andrew H Fischer, Kenneth A Jacobson, Jack Rose, and Rolf Zeller. Hematoxylin and eosin staining of tissue
and cell sections. Cold Spring Harbor Protocols , 2008(5):pdb–prot4986, 2008.
[6] S. Tripathi, S. Mishra, and S. K. Singh. Routine colon cancer detection using local image descriptors. In 2016
IEEE Region 10 Conference (TENCON) , pages 2062–2065, Nov 2016.
[7] Daniele Zink, Andrew H Fischer, and Jeffrey A Nickerson. Nuclear structure in cancer cells. Nature reviews
cancer , 4(9):677, 2004.
[8] Humayun Irshad, Antoine Veillard, Ludovic Roux, and Daniel Racoceanu. Methods for nuclei detection, seg-
mentation, and classification in digital histopathology: a review—current status and future potential. IEEE
reviews in biomedical engineering , 7:97–114, 2014.
[9] H Peter Rodemann and Hans-Oliver Rennekampff. Functional diversity of fibroblasts. In Tumor-Associated
Fibroblasts and their Matrix , pages 23–36. Springer, 2011.
[10] Pin Wang, Xianling Hu, Yongming Li, Qianqian Liu, and Xinjian Zhu. Automatic cell nuclei segmentation and
classification of breast cancer histopathology images. Signal Processing , 122:1–13, 2016.
[11] James Diamond, Neil H Anderson, Peter H Bartels, Rodolfo Montironi, and Peter W Hamilton. The use of mor-
phological characteristics and texture analysis in the identification of tissue composition in prostatic neoplasia.
Human pathology , 35(9):1121–1131, 2004.
[12] Scott Doyle, Mark Hwang, Kinsuk Shah, Anant Madabhushi, Michael Feldman, and John Tomaszeweski. Auto-
mated grading of prostate cancer using architectural and textural image features. In 2007 4th IEEE International
Symposium on Biomedical Imaging: From Nano to Macro , pages 1284–1287. IEEE, 2007.
[13] Metin N Gurcan, Laura Boucheron, Ali Can, Anant Madabhushi, Nasir Rajpoot, and Bulent Yener. Histopatho-
logical image analysis: A review. IEEE reviews in biomedical engineering , 2:147, 2009.
16APREPRINT - FEBRUARY 22, 2022
[14] Laura E Boucheron. Object-and spatial-level quantitative analysis of multispectral histopathology images for
detection and characterization of cancer . University of California at Santa Barbara, 2008.
[15] Korsuk Sirinukunwattana, David RJ Snead, and Nasir M Rajpoot. A novel texture descriptor for detection of
glandular structures in colon histology images. In Medical Imaging 2015: Digital Pathology , volume 9420, page
94200S. International Society for Optics and Photonics, 2015.
[16] Jun Shi, Jinjie Wu, Yan Li, Qi Zhang, and Shihui Ying. Histopathological image classification with color pat-
tern random binary hashing-based pcanet and matrix-form classifier. IEEE journal of biomedical and health
informatics , 21(5):1327–1337, 2017.
[17] Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. nature , 521(7553):436, 2015.
[18] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan,
Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Proceedings of the IEEE
conference on computer vision and pattern recognition , pages 1–9, 2015.
[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In
Proceedings of the IEEE conference on computer vision and pattern recognition , pages 770–778, 2016.
[20] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition.
arXiv preprint arXiv:1409.1556 , 2014.
[21] Babak Ehteshami Bejnordi, Mitko Veta, Paul Johannes Van Diest, Bram Van Ginneken, Nico Karssemeijer,
Geert Litjens, Jeroen AWM Van Der Laak, Meyke Hermsen, Quirine F Manson, Maschenka Balkenhol, et al.
Diagnostic assessment of deep learning algorithms for detection of lymph node metastases in women with breast
cancer. Jama , 318(22):2199–2210, 2017.
[22] Jianpeng Zhang, Yong Xia, Yutong Xie, Michael Fulham, and David Dagan Feng. Classification of medical
images in the biomedical literature by jointly using deep and handcrafted visual features. IEEE journal of
biomedical and health informatics , 22(5):1521–1530, 2018.
[23] Haibo Wang, Angel Cruz Roa, Ajay N Basavanhally, Hannah L Gilmore, Natalie Shih, Mike Feldman, John
Tomaszewski, Fabio Gonzalez, and Anant Madabhushi. Mitosis detection in breast cancer pathology images by
combining handcrafted and convolutional neural network features. Journal of Medical Imaging , 1(3):034003,
2014.
[24] Dan C Cires ¸an, Alessandro Giusti, Luca M Gambardella, and J ¨urgen Schmidhuber. Mitosis detection in breast
cancer histology images with deep neural networks. In International Conference on Medical Image Computing
and Computer-assisted Intervention , pages 411–418. Springer, 2013.
[25] Sajjad Muhammad, Khan Salman, Khan Muhammad, Wanqing Wu, Amin Ullah, and Sung Wook Baik. Multi-
grade brain tumor classification using deep cnn with extensive data augmentation. Journal of computational
science , 30:174–182, 2019.
[26] Shui-Hua Wang, Khan Muhammad, Jin Hong, Arun Kumar Sangaiah, and Yu-Dong Zhang. Alcoholism iden-
tification via convolutional neural network based on parametric relu, dropout, and batch normalization. Neural
Computing and Applications , Dec 2018.
[27] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional
networks. In Proceedings of the IEEE conference on computer vision and pattern recognition , pages 4700–4708,
2017.
[28] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna. Rethinking the
inception architecture for computer vision. arxiv 2015. arXiv preprint arXiv:1512.00567 , 1512, 2015.
[29] Daniele Rav `ı, Charence Wong, Fani Deligianni, Melissa Berthelot, Javier Andreu-Perez, Benny Lo, and Guang-
Zhong Yang. Deep learning for health informatics. IEEE journal of biomedical and health informatics , 21(1):4–
21, 2017.
[30] Yangqing Jia, Evan Shelhamer, Jeff Donahue, Sergey Karayev, Jonathan Long, Ross Girshick, Sergio Guadar-
rama, and Trevor Darrell. Caffe: Convolutional architecture for fast feature embedding. In Proceedings of the
22nd ACM international conference on Multimedia , pages 675–678. ACM, 2014.
[31] Ken Chatfield, Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. Return of the devil in the details:
Delving deep into convolutional nets. arXiv preprint arXiv:1405.3531 , 2014.
[32] Korsuk Sirinukunwattana, Shan e Ahmed Raza, Yee-Wah Tsang, David RJ Snead, Ian A Cree, and Nasir M Ra-
jpoot. Locality sensitive deep learning for detection and classification of nuclei in routine colon cancer histology
images. IEEE Trans. Med. Imaging , 35(5):1196–1206, 2016.
17APREPRINT - FEBRUARY 22, 2022
[33] Hang Chang, Leandro A Loss, and Bahram Parvin. Nuclear segmentation in h&e sections via multi-reference
graph cut (mrgc). In International symposium biomedical imaging , 2012.
[34] Robert M Haralick, Karthikeyan Shanmugam, et al. Textural features for image classification. IEEE Transactions
on systems, man, and cybernetics , (6):610–621, 1973.
[35] Yann LeCun, Bernhard Boser, John S Denker, Donnie Henderson, Richard E Howard, Wayne Hubbard, and
Lawrence D Jackel. Backpropagation applied to handwritten zip code recognition. Neural computation ,
1(4):541–551, 1989.
[36] Andrej Karpathy. Stanford university cs231n: convolutional neural networks for visual recognition. URL:
http://cs231n. stanford. edu/syllabus. html , 2018.
[37] Jeff Donahue, Yangqing Jia, Oriol Vinyals, Judy Hoffman, Ning Zhang, Eric Tzeng, and Trevor Darrell. Decaf:
A deep convolutional activation feature for generic visual recognition. In International conference on machine
learning , pages 647–655, 2014.
[38] Thomas P V ogl, JK Mangis, AK Rigler, WT Zink, and DL Alkon. Accelerating the convergence of the back-
propagation method. Biological cybernetics , 59(4-5):257–263, 1988.
[39] Christopher M Bishop. Pattern recognition and machine learning . springer, 2006.
[40] Kevin P Murphy. Machine learning: a probabilistic perspective . MIT press, 2012.
[41] Mateusz Buda, Atsuto Maki, and Maciej A Mazurowski. A systematic study of the class imbalance problem in
convolutional neural networks. Neural Networks , 106:249–259, 2018.
[42] Haibo He, Yang Bai, Edwardo A Garcia, and Shutao Li. Adasyn: Adaptive synthetic sampling approach for
imbalanced learning. In 2008 IEEE International Joint Conference on Neural Networks (IEEE World Congress
on Computational Intelligence) , pages 1322–1328. IEEE, 2008.
[43] L Torrey and J Shavlik. Transfer learning. handbook of research on machine learning applications, vol. 3, 2009.
[44] Yinyin Yuan, Henrik Failmezger, Oscar M Rueda, H Raza Ali, Stefan Gr ¨af, Suet-Feung Chin, Roland F Schwarz,
Christina Curtis, Mark J Dunning, Helen Bardwell, et al. Quantitative image analysis of cellular heterogeneity
in breast tumors complements genomic profiling. Science translational medicine , 4(157):157ra143–157ra143,
2012.
18