Knowledge distillation from multi-modal to
mono-modal segmentation networks
Minhao Hu1;2?, Matthis Maillard2?(), Ya Zhang1(), Tommaso Ciceri2,
Giammarco La Barbera2, Isabelle Bloch2, and Pietro Gori2
1CMIC, Shanghai Jiao Tong University, Shanghai, China
2LTCI, T el ecom Paris, Institut Polytechnique de Paris, France
matthis.maillard@telecom-paris.fr
yazhang@sjtu.edu.cn
Abstract. The joint use of multiple imaging modalities for medical im-
age segmentation has been widely studied in recent years. The fusion of
information from dierent modalities has demonstrated to improve the
segmentation accuracy, with respect to mono-modal segmentations, in
several applications. However, acquiring multiple modalities is usually
not possible in a clinical setting due to a limited number of physicians
and scanners, and to limit costs and scan time. Most of the time, only
one modality is acquired. In this paper, we propose KD-Net, a framework
to transfer knowledge from a trained multi-modal network (teacher) to
a mono-modal one (student). The proposed method is an adaptation
of the generalized distillation framework where the student network is
trained on a subset (1 modality) of the teacher's inputs (n modalities).
We illustrate the eectiveness of the proposed framework in brain tumor
segmentation with the BraTS 2018 dataset. Using dierent architectures,
we show that the student network eectively learns from the teacher and
always outperforms the baseline mono-modal network in terms of seg-
mentation accuracy.
1 Introduction
Using multiple modalities to automatically segment medical images has become
a common practice in several applications, such as brain tumor segmentation [11]
or ischemic stroke lesion segmentation [10]. Since dierent image modalities can
accentuate and better describe dierent tissues, their fusion can improve the seg-
mentation accuracy. Although multi-modal models usually give the best results,
it is often dicult to obtain multiple modalities in a clinical setting due to a
limited number of physicians and scanners, and to limit costs and scan time. In
many cases, especially for patients with pathologies or for emergency, only one
modality is acquired.
Two main strategies have been proposed in the literature to deal with prob-
lems where multiple modalities are available at training time but some or most
?The two rst authors contributed equally to this paper.arXiv:2106.09564v1  [cs.CV]  17 Jun 20212 M.Hu et al.
of them are missing at inference time. The rst one is to train a generative model
to synthesize the missing modalities and then perform multi-modal segmenta-
tion. In [13], the authors have shown that using a synthesized modality helps
improving the accuracy of classication of brain tumors. Ben Cohen et al. [1]
generated PET images from CT scans to reduce the number of false positives in
the detection of malignant lesions in livers. Generating a synthesized modality
has also been shown to improve the quality of the segmentation of white matter
hypointensities [12]. The main drawback of this strategy is that it is compu-
tationally cumbersome, especially when many modalities are missing. In fact,
one needs to train one generative network per missing modality in addition to a
multi-modal segmentation network.
The second strategy consists in learning a modality-invariant feature space
that encodes the multi-modal information during training, and that allows for all
possible combinations of modalities during inference. Within this second strat-
egy, Havaei et al. proposed HeMIS [4], a model that, for each modality, trains a
dierent feature extractor. The rst two moments of the feature maps are then
computed and concatenated in the latent space from which a decoder is trained
to predict the segmentation map. Dorent et al. [3], inspired by HeMIS, proposed
U-HVED where they introduced skip-connections by considering intermediate
layers, before each down-sampling step, as a feature map. This network outper-
formed HeMIS on BraTS 2018 dataset. In [2], instead of fusing the layers by
computing mean and variance, the authors learned a mapping function from the
multiple feature maps to the latent space. They claimed that computing the
moments to fuse the maps is not satisfactory since it makes each modality con-
tribute equally to the nal result which is inconsistent with the fact that each
modality highlights dierent zones. They obtained better results than HeMIS
on BraTS 2015 dataset. This second strategy has good results only when one
or two modalities are missing, however, when only one modality is available, it
has worse results than a model trained on this specic modality. This kind of
methods is therefore not suitable for a clinical setting where only one modality
is usually acquired, such as pre-operative neurosurgery or radiotherapy.
In this paper, in contrast to the previously presented methods, we propose a
framework to transfer knowledge from a multi-modal network to a mono-modal
one. The proposed method is based on generalized knowledge distillation [9]
which is a combination of distillation [5] and privileged information [14]. Distil-
lation has originally been designed for classication problems to make a small
network (Student) learn from an ensemble of networks or from a large network
(Teacher). It has been applied to image segmentation in [8,15] where the same in-
put modalities have been used for the Teacher network and the Student network.
In [15], the Student learns from the Teacher only thanks to a loss term between
their outputs. In [8], the authors also constrained the intermediate layers of the
Student to be similar to the ones of the Teacher. With a dierent perspective,
the framework of privileged information was designed to boost the performance
of a Student model by learning from both the training data and a Teacher model
with privileged and additional information. In generalized knowledge distillation,KD-Net 3
one uses distillation to extract useful knowledge from the privileged information
of the Teacher [9]. In our case, Teacher and Student have the same architec-
ture (i.e. same number of parameters) but the Teacher can learn from multiple
input modalities (additional information) whereas the Student from only one.
The proposed framework is based on two encoder-decoder networks, which have
demonstrated to work well in image segmentation [7], one for the Student and
one for the Teacher. Importantly, the proposed framework is generic since it
works for dierent architectures of the encoder-decoder networks. Each encoder
summarizes its input space to a latent representation that captures important
information for the segmentation. Since the Teacher and the Student process
dierent inputs but aim at extracting the same information, we make the as-
sumption that their rst layers should be dierent, whereas the last layers and
especially the latent representations (i.e. bottleneck) should be similar. By forc-
ing the latent space of the Student to resemble the one of the Teacher, we make
the hypothesis that the Student should learn from the additional information
of the Teacher. To the best of our knowledge, this is the rst time that the
generalized knowledge distillation strategy is adapted to guide the learning of
a mono-modal student network using a multi-modal teacher network. We show
the eectiveness of the proposed method using the BraTS 2018 dataset [11] for
brain tumor segmentation.
The paper is organized as follows. First, we present the proposed framework,
called KD-Net and illustrated in Figure 1, and how the Student learns from the
Teacher and the reference segmentation. Then, we present the implementation
details and the results on the BraTS 2018 dataset [11].
KD loss GT loss KL loss
128×128×128×1
128×128×128×4
MaxPool3d Trilinear interpolation Softmax Conv3d InstanceNorm3d LeakyReLUReference 
segmentation
Teacher
Student
Fig. 1. Illustration of the proposed framework. Both Teacher and Student have the
same architecture adapted from nnUNet [7]. First, the Teacher is trained using only
the reference segmentation (GT loss). Then, the student network is trained using all
proposed losses: KL loss, KD loss and GT loss.4 M.Hu et al.
2 KD-Net
The goal of the proposed framework is to train a mono-modal segmentation
network (Student) by leveraging the knowledge from a well-trained multi-modal
segmentation network (Teacher). Except for the number of input channels, both
networks have the same encoder-decoder architecture with skip connections. The
multi-modal input xi=fxi
n;n= 1:::Ngis the concatenation of the Nmodalities
for theithsample of the dataset. Let EtandDt(resp.EsandDs) denote the
encoder and decoder parts of the Teacher (resp. Student). The Teacher network
ft(xi) =DtEt(xi) receives as input multiple modalities whereas the student
networkfs(xi
k) =DsEs(xi
k) only one modality xi
k,kbeing a xed integer
between 1 and N.
We rst train the Teacher, using only the reference segmentation as target.
Then, we train the Student using three dierent losses: the knowledge distillation
term, the dissimilarity between the latent spaces, and the reference segmentation
loss. Note that the weights of the Teacher are frozen during the training of the
Student and the error of the Student is not back-propagated to the Teacher.
The rst two terms allow the Student to learn from the Teacher by using the
soft prediction of the latter as target and by forcing the encoded information
(i.e. bottleneck) of the Student to be similar to the one of the Teacher. The last
term makes the predicted segmentation of the Student similar to the reference
segmentation.
2.1 Generalized knowledge distillation
Following the strategy of generalized knowledge distillation [9], we transfer useful
knowledge from the additional information of the Teacher to the Student using
the soft label targets of the Teacher. These are computed as follows:
si=(ft(xi)=T) (1)
whereis the softmax function and T, the temperature parameter, is a strictly
positive value. The parameter Tcontrols the softness of the target, and the higher
it is, the softer the target. The idea of using soft targets is to uncover relations
between classes that would be harder to detect with hard labels. The eectiveness
of using a temperature parameter to soften the labels was demonstrated in [5].
The knowledge distillation loss is dened as:
LKD=X
i
(1 Dice (si;(fs(xi
k)))) +BCE (s
i;(fs(xi
k))
(2)
whereDice is the Dice score, BCE the binary cross-entropy measure and s
i
the binary prediction of the teacher. We need to binarize sisince the soft labels
cannot be used in the binary cross-entropy. The dice score ( Dice ) measures the
similarity of the shape of two ensembles. Hence, it globally measures how the
Teacher and Student's segmentation maps are close to each other. By contrast,
the binary cross-entropy ( BCE ) is computed for each pixel independently andKD-Net 5
therefore it is a local measure. We use the combination of these two terms to
globally and locally measure the distance between the Student prediction and
the Teacher soft labels.
2.2 Latent space
We speculate that Teacher and Student, having dierent inputs, should also
encode dierently the information in the rst layers, the ones related to low-
level image properties, such as color, texture and edges. By contrast, the deepest
layers closer to the bottleneck, and related to higher level properties, should be
more similar. Furthermore, we make the assumption that an encoder-decoder
network encodes the information to correctly segment the input images in its
latent space. Based on that, we propose to force the Student to learn from the
additional information of the Teacher encoded in its bottleneck (and partially in
the deepest layers) by making their latent representations as close as possible.
To this end, we apply the Kullback-Leibler (KL) divergence as a loss function
between the teacher and student's bottlenecks:
LKL(p;q) =X
iX
jqi(j) logqi(j)
pi(j)
(3)
wherepi(resp.qi) are the attened and normalized vector of the bottleneck
Es(xi
k) (respEt(xi)). Note that this function is not symmetric and we put the
vectors in that order because we want the distribution of the Student's bottleneck
to be similar to the one of the Teacher.
2.3 Objective function
We add a third term to the objective function to make the predicted segmen-
tation as close as possible to the reference segmentation. It is the sum of the
Dice loss (Dice ) and the binary cross-entropy ( BCE ) for the same reasons as in
Section 2.1. We call it LGT:
LGT=X
i
(1 Dice (yi;(fs(xi
k)))) +BCE (yi;(fs(xi
k))
: (4)
whereyidenotes the reference segmentation of the ithsample in the dataset.
The complete objective function is then:
L=LKD+ (1 )LGT+LKL (5)
with2[0;1] and2R+. The imitation parameter balances the inuence
of the reference segmentation with the one of the Teacher's soft labels. The
greater the value, the greater the inuence of the Teacher's soft labels. The
parameter is instead needed to balance the magnitude of the KL loss with
respect to the other two losses.6 M.Hu et al.
3 Results and Discussion
3.1 Dataset
We evaluate the performance of the proposed framework on a publicly avail-
able dataset from the BraTS 2018 Challenge [11]. It contains MR scans from
285 patients with four modalities: T1, T2, T1 contrasted-enhanced (T1ce) and
Flair. The goal of the challenge is to segment three sub-regions of brain tumors:
whole tumor (WT), tumor core (TC) and enhancing tumor (ET). We apply a
central crop of size 128 128128 and a random ip along each axis for data
augmentation. For each modality, only non-zero voxels have been normalized by
subtracting the mean and dividing by standard deviation. Due to memory and
time constraint, we subsample the images to the size 64 6464.
3.2 Implementation details
We adopt the encoder-decoder architecture described in Figure 1. Empirically,
we found that the best parameters for the objective function are = 0:75,T= 5
and= 10. We used Adam optimizer for 500 epochs with a learning rate equal
to 0.0001 that is multiplied by 0.2 when the validation loss has not decreased
for 50 epochs. We run a three fold cross validation on the 285 training cases
of BraTS 2018. The training of the baseline, the Teacher or the Student takes
approximately 12 hours on a NVIDIA P100 GPU.
3.3 Results
In our experiments, the Teacher uses all four modalities (T1, T2, T1ce and
Flair concatenated) and the Student uses only T1ce. We choose T1ce for the
Student since this is the standard modality used in pre-operative neurosurgery
or radiotherapy.
Model comparison: To demonstrate the eectiveness of the proposed frame-
work, we rst compare it to a baseline model. Its architecture is the same as the
encoder-decoder network in Figure 1 and it is trained using only the T1ce modal-
ity as input. We also compare it to two other models, U-HVED and HeMIS, using
only T1ce as input. Results were directly taken from [3]. The results are visible
in Table 1. Our method outperforms U-HVED and HeMIS in the segmentation
of all three tumor components. KD-Net also seems to obtain better results than
the method proposed in [2] (again when using only T1ce as input). The authors
show results on the BraTS 2015 dataset and therefore they are not directly
comparable to KD-Net. Furthermore, we could not nd online their code. Nev-
ertheless, the results of HeMIS [4] on BraTS 2015 (in [2]) and on BraTS 2018
(in [3]) suggest that the observations of BraTS 2018 seem to be more dicult
to segment. Since the method proposed in [2] has worst results than ours on a
dataset that seems easier to segment, this should also be the case for the BraTS
2018 dataset. However, this should be conrmed.KD-Net 7
Table 1. Comparison of 3 models using the dice score on the tumor regions. Results
of U-HVED and HeMIS are taken from the article [3], where the standard deviations
were not provided.
Model ET TC WT
Baseline (nnUnet [7]) 68:11:27 80 :282:44 77 :061:47
Teacher (4 modalities) 69:471:86 80 :771:18 88 :480:79
U-HVED 65:5 66 :7 62 :4
HeMIS 60:8 58 :5 58 :5
Ours 71 :671:22 81 :451:25 76:981:54
Ablation study: To evaluate the contribution of each loss term, we did an
ablation study by removing each term from the objective function dened in
Eq. 5. Table 2 shows the results using either 0 or 4 skip-connections both in the
Student and Teacher networks. We observe that both the KL and KD loss im-
proves the results with respect to the baseline model, especially for the enhanced
tumor and tumor core. This also demonstrates that the proposed framework is
generic and it works with dierent encoder-decoder architectures. More results
can be found in the supplementary material.
Table 2. Ablation study of the loss terms. We compare the results of the model
trained with 3 dierent objective functions: the baseline using only the GT loss, KD-
Net trained with only the KL term and KD-Net with the complete objective function.
We also tested it with 0 or 4 skip-connections for both the Student and the Teacher.
Skip
connectionsModel Loss ET TC WT
4 Baseline GT 68:11:27 80:282:44 77:061:47
4 Teacher GT 69:471:86 80:771:18 88:480:79
4 KD-Net GT+KL 70:001:51 80:851:82 77 :081:29
4 KD-Net GT+KD 69:221:19 80:541:66 76:831:36
4 KD-Net GT+KL+KD 71 :671:2281 :451:25 76:981:54
0 Baseline GT 42:953:42 69:441:37 69:411:52
0 Teacher GT 42:592:54 69:791:63 75:930:33
0 KD-Net GT+KL 47 :590:98 70:961:73 71:411:2
0 KD-Net GT+KD 44:81:1 70:122:42 70:191:4
0 KD-Net GT+KL+KD 46:232:91 70:732:47 71 :931:26
Qualitative results: In Figure 2, we show some qualitative results of the
proposed framework and compare them with the ones obtained using the base-
line method. We can see that the proposed framework allows the Student to8 M.Hu et al.
discard some outliers and predict segmentation labels of higher quality. In the
experiments, the student uses as input only T1ce, which clearly highlights the
enhancing tumor. Remarkably, it seems that the Student learns more in this
region (see Figure 2 and Table 1). The knowledge distilled from the Teacher
seems to help the Student learn more where it is supposed to be \stronger".
More qualitative results can be found in the supplementary material.
Fig. 2. Qualitative results obtained using the the baseline and the proposed framework
(Student). We show the slice of a subject with the corresponding 3 segmentation labels.
Observations: It is important to remark that we also tried to expand the
Student network by rst synthesizing another modality, such as the Flair, from
the T1ce and then using it, together with the T1ce, for segmenting the tumor
labels. Results were actually worse than the baseline and the computational
time quite prohibitive. We also tried sharing the weights between the Teacher
and the Student in the deepest layers of the networks to help transferring the
knowledge. The intuition behind it was that since the bottlenecks should be the
same, the information in the deepest layers should be handled identically. The
results were almost identical, but slightly worse, to the ones obtained with the
proposed framework presented in Figure 1. In this paper, we used the nnUNet[7]
as network for the Student and Teacher, but theoretically any other encoder-
decoder architecture, such as the one in [6], could be used.KD-Net 9
4 Conclusions
We present a novel framework to transfer knowledge from a multi-modal segmen-
tation network to a mono-modal one. To this end, we propose to use a twofold
approach. We employ the strategy of generalized knowledge distillation and, in
addition, we also constrain the latent representation of the Student to be similar
to the one of the Teacher. We validate our method in brain tumor segmen-
tation, achieving better results than state-of-the-art methods when using only
T1ce on Brats 2018. The proposed framework is generic and can be applied to
any encoder-decoder segmentation network. The gain in segmentation accuracy
and robustness to errors produced by the proposed framework makes it highly
valuable for real-world clinical scenarios where only one modality is available at
test time.
5 Acknowledgment
M.Hu is grateful for nancial support from China Scholarship Council.This work
is supported by SHEITC (No. 2018-RGZN-02046), 111 plan (No. BP0719010),
and STCSM (No. 18DZ2270700). M. Maillard was supported by a grant of IMT,
Fondation Mines-T el ecom and Institut Carnot TSN, through the \Futur & Rup-
tures" program.
References
1. Ben-Cohen, A., Klang, E., Raskin, S., Soer, S., Ben-Haim, S., Konen, E., Amitai,
M., Greenspan, H.: Cross-modality synthesis from CT to PET using FCN and
GAN networks for improved automated lesion detection. Engineering Applications
of Articial Intelligence 78, 186{194 (2018)
2. Chen, C., Dou, Q., Jin, Y., Chen, H., Qin, J., Heng, P.A.: Robust Multimodal
Brain Tumor Segmentation via Feature Disentanglement and Gated Fusion. In:
MICCAI. vol. LNCS 11766, pp. 447{456. Springer, Cham (2019)
3. Dorent, R., Joutard, S., Modat, M., Ourselin, S., Vercauteren, T.: Hetero-Modal
Variational Encoder-Decoder for Joint Modality Completion and Segmentation.
In: MICCAI. vol. LNCS 11765, pp. 74{82. Springer (2019)
4. Havaei, M., Guizard, N., Chapados, N., Bengio, Y.: HeMIS: Hetero-Modal Image
Segmentation. In: MICCAI. vol. LNCS 9901, pp. 469{477. Springer (2016)
5. Hinton, G., Vinyals, O., Dean, J.: Distilling the Knowledge in a Neural Network.
Deep Learning and Representation Learning Workshop: NIPS 2015 (2015)
6. Ibtehaz, N., Rahman, M.S.: MultiResUNet : Rethinking the U-Net Architecture for
Multimodal Biomedical Image Segmentation. Neural Networks 121, 74{87 (2020)
7. Isensee, F., Kickingereder, P., Wick, W., Bendszus, M., Maier-Hein, K.H.: No New-
Net. In: BrainLes - MICCAI Workshop. vol. LNCS 11384, pp. 234{244. Springer
(2019)
8. Liu, Y., Chen, K., Liu, C., Qin, Z., Luo, Z., Wang, J.: Structured Knowledge
Distillation for Semantic Segmentation. In: CVPR. pp. 2604{2613 (2019)
9. Lopez-Paz, D., Bottou, L., Sch olkopf, B., Vapnik, V.: Unifying distillation and
privileged information. In: ICLR (2016)10 M.Hu et al.
10. Maier, O., et al.: ISLES 2015 - A public evaluation benchmark for ischemic stroke
lesion segmentation from multispectral MRI. Medical Image Analysis 35, 250{269
(2017)
11. Menze, B.H., et al.: The multimodal brain tumor image segmentation benchmark
(BRATS). IEEE Transactions on Medical Imaging 34(10), 1993{2024 (2015)
12. Orbes-Arteaga, M., Cardoso, M.J., Srensen, L., Modat, M., Ourselin, S., Nielsen,
M., Pai, A.: Simultaneous synthesis of FLAIR and segmentation of white matter
hypointensities from T1 MRIs. In: MIDL (2018)
13. van Tulder, G., de Bruijne, M.: Why Does Synthesized Data Improve Multi-
sequence Classication? In: MICCAI. vol. LNCS 9349, pp. 531{538. Springer,
Cham (2015)
14. Vapnik, V., Izmailov, R.: Learning using privileged information: Similarity control
and knowledge transfer. Journal of Machine Learning Research 16(61), 2023{2049
(2015)
15. Xie, J., Shuai, B., Hu, J.F., Lin, J., Zheng, W.S.: Improving Fast Segmentation
With Teacher-student Learning. In: British Machine Vision Conference (BMVC)
(2018)