Deep Supervised Information Bottleneck Hashing for Cross-modal Retrieval
based Computer-aided Diagnosis
Yufeng Shi1, Shuhuang Chen1, Xinge You1*, Qinmu Peng1, Weihua Ou2, Yue Zhao3
1Huazhong University of Science and Technology
2Guizhou Normal University
3Hubei University
fyufengshi17, shuhuangchen, youxg, pengqinmu g@hust.edu.cn, ouweihuahust@gmail.com, zhaoyhu@hubu.edu.cn
Abstract
Mapping X-ray images, radiology reports, and other medi-
cal data as binary codes in the common space, which can
assist clinicians to retrieve pathology-related data from het-
erogeneous modalities (i.e., hashing-based cross-modal med-
ical data retrieval), provides a new view to promot computer-
aided diagnosis. Nevertheless, there remains a barrier to boost
medical retrieval accuracy: how to reveal the ambiguous se-
mantics of medical data without the distraction of superflu-
ous information. To circumvent this drawback, we propose
Deep Supervised Information Bottleneck Hashing (DSIBH),
which effectively strengthens the discriminability of hash
codes. Specifically, the Deep Deterministic Information Bot-
tleneck (Yu, Yu, and Pr ´ıncipe 2021) for single modality is
extended to the cross-modal scenario. Benefiting from this,
the superfluous information is reduced, which facilitates the
discriminability of hash codes. Experimental results demon-
strate the superior accuracy of the proposed DSIBH com-
pared with state-of-the-arts in cross-modal medical data re-
trieval tasks.
The rapid development of medical technology not only
provides diverse medical examinations but also produces
tremendous amounts of medical data, ranging from X-ray
images to radiology reports. It is an experience-demanding,
time-consuming, and error-prone job to manually assess
medical data and diagnose disease. To reduce the work
burden of physicians and optimize the diagnostic process,
computer-aided diagnosis (CAD) systems including classi-
fier based CAD(Shi et al. 2020; In ´es et al. 2021) and content-
based image retrieval (CBIR) based CAD(Yang et al. 2020;
Fang, Fu, and Liu 2021) have been designed to automatically
identify illness. Although the two types of methods greatly
promote the development of CAD, existing systems ignore
the character of current medical data, which is diverse in
modality and huge in terms of scale. Therefore, we introduce
cross-modal retrieval (CMR) (Wang et al. 2016) techniques
and construct a CMR-based CAD method using semantic
hashing (Wang et al. 2017) to handle the above challenges.
With the help of CMR that projects multimodal data into
the common space, samples from different modalities can be
directly matched without the interference of heterogeneity.
*Contact Author
Accepted by the AAAI-22 Workshop on Information Theory for
Deep Learning (IT4DL).Therefore, CMR-based CAD can not only retrieve the se-
mantically similar clinical profiles in heterogeneous modal-
ities but also provide diagnosis results according to the pre-
vious medical advice. Compared with the classifier-based
CAD that only provides diagnosis results, CMR-based CAD
is more acceptable due to the interpretability brought by
retrieved profiles. Compared with the CBIR-based CAD,
CMR-based CAD wins on its extended sight of multi-modal
data, which meets the requirement of current medical data.
Recently, extensive work on hashing-based CMR that
maps data from different modalities into the same hamming
space, has been done by researchers to achieve CMR (Li
et al. 2018; Zhu et al. 2020; Yu et al. 2021). Due to its com-
pact binary codes and XOR distance calculation, hashing-
based CMR possesses low memory usage and high query
speed (Wang et al. 2017), which is also compatible with
the huge volume of current medical data. In terms of ac-
curacy, the suitable hashing-based solutions for CMR-based
CAD are deep supervised hashing (DSH) methods (Xie et al.
2020; Zhan et al. 2020; Yao et al. 2021). With the guid-
ance of manual annotations, deep supervised methods usu-
ally perform hash code learning based on the original data
via neural networks. Inspired by the information bottle-
neck principle (Tishby, Pereira, and Bialek 1999), the above-
mentioned optimization procedure can be viewed as build-
ing hash code Gabout a semantic label Ythrough samples
in different modalities X=/braceleftbig
X1,X2/bracerightbig
, which can be for-
mulated as:
maxL=I(G;Y)−βI(G;X), (1)
whereI(·;·)represents the mutual information, and βis a
hyper-parameter. As quantified by I(G;Y), current DSH
methods model the semantic annotations to establish pair-
wise, triplet-wise or class-wise relations, and maximize the
correlation between hash codes and the semantic relations.
Despite the consideration of semantic relations, the neglect
ofI(G;X)will result in the retention of redundant infor-
mation in the original data, thus limiting the improvement
of the retrieval accuracy. I(G;X)measures the correlation
between the hash code Gand the data from two modalities
X, which can be used to reduce the superfluous informa-
tion from medical data, and constrain the hash code to grasp
the correct semantics from annotations. Therefore, it can be
expected that the optimization of Eq. (1) can strengthen thearXiv:2205.08365v1  [cs.LG]  6 May 2022discriminability of hash codes, which improves the accuracy
of CMR-based CAD.
To perform CMR-based CAD, we design a novel method
named Deep Supervised Information Bottleneck Hash-
ing (DSIBH), which optimizes the information bottleneck
to strengthen the discriminability of hash codes. Specifi-
cally, to avoid variational inference and distribution assump-
tion, we extend the Deep Deterministic Information Bot-
tleneck (DDIB) (Yu, Yu, and Pr ´ıncipe 2021) from single
modality to the cross-modal scenario for hash code learning.
To summarize, our main contributions are fourfold:
• The cross-modal retrieval technique based on semantic
hashing is introduced to establish computer-aided diag-
nosis systems, which is suitable for the current large-
scale multi-modal medical data.
• A deep hashing method named DSIBH, which optimizes
the hash code learning procedure following the informa-
tion bottleneck principle to reduce the distraction of su-
perfluous information, is proposed for CMR-based CAD.
• To reduce the adverse impact of variational inference
or distribution assumption, the Deep Deterministic In-
formation Bottleneck is elegantly extended to the cross-
modal scenario for hash code learning.
• Experiments on the large-scale multi-modal medical
dataset MIMIC-CXR show that DSIBH can strengthen
the discriminability of hash codes more effectively than
other methods, thus boosting the retrieval accuracy.
Related Work
In this section, representative CAD approaches and hashing-
based solutions of cross-modal retrieval are briefly reviewed.
To make readers easier to understand our work, some knowl-
edge of the DDIB is also introduced.
Computer-aided Diagnosis
CAD approaches generally fall into two types including
classifier-based CAD and CBIR-based CAD. Thanks to the
rapid progress of deep learning, classifier-based CAD meth-
ods (Zhang et al. 2019; de La Torre, Valls, and Puig 2020)
can construct task-specific neural networks to categorize
histopathology images and employ the outcomes as the diag-
nosis. On the other side, CBIR-based CAD can provide clin-
ical evidence since they retrieve and visualize images with
the most similar morphological profiles. According to the
data type of representations, existing CBIR methods can be
divided into continuous value CBIR (Erfankhah et al. 2019;
Zhen et al. 2020) and hashing-based CBIR (Hu et al. 2020;
Yang et al. 2020). In the age of big data, the latter increas-
ingly become mainstream due to the low memory usage and
high query speed brought by hashing. Although substantial
efforts have been made to analyse clinical image, medical
data such as radiology reports in other modalities are ig-
nored. Consequently, CAD is restricted in single modality
and the cross-modal relevance between different modalities
still waits to be explored.Cross-modal Retrieval
Cross-modal hashing has made remarkable progress in han-
dling cross-modal retrieval, and this kind of methods can
be roughly divided into two major types including unsuper-
vised approaches and supervised approaches in terms of the
consideration of semantic information. Due to the absence of
semantic information, the former usually relies on data dis-
tributions to align semantic similarities of different modali-
ties (Liu et al. 2020; Yu et al. 2021). For example, Collec-
tive Matrix Factorization Hashing (Ding et al. 2016) learns
unified hash codes by collective matrix factorization with
a latent factor model to capture instance-level correlations.
Recently, Deep Graph-neighbor Coherence Preserving Net-
work (Yu et al. 2021) extra explores graph-neighbor coher-
ence to describe the complex data relationships. Although
data distributions indeed help to solve cross-modal retrieval
to some extent, one should note that unsupervised methods
fail to manage the high-level semantic relations due to the
neglect of manual annotations.
Supervised hashing methods are thereafter proposed to
perform hash code learning with the guidance of manual
annotations. Data points are encoded to express semantic
similarity such as pair-wise(Shen et al. 2017; Wang et al.
2019), triplet-wise (Hu et al. 2019; Song et al. 2021) or
multi-wise similarity relations(Cao et al. 2017; Li et al.
2018). As an early attempt with deep learning, Deep Cross-
modal Hashing (Jiang and Li 2017) directly encodes origi-
nal data points by minimizing the negative log likelihood of
the cross-modal similarities. To discover high-level semantic
information, Self-Supervised Adversarial Hashing (Li et al.
2018) harnesses a self-supervised semantic network to pre-
serve the pair-wise relationships. Although various relations
have been built between the hash code and the semantic la-
bels, the aforementioned algorithms still suffer from the dis-
traction of superfluous information, which is caused by the
connections between the hash code and the original data,
Consequently, for CMR-based CAD, there remains a need
for a deep hashing method which can reduce the superfluous
information to strengthen the discriminability of hash codes.
Deep Deterministic Information Bottleneck
Despite great efforts to handle the ambiguous semantics of
medical data, the discriminability of hash codes still needs
to be strengthened. To alleviate such limitation, a promising
solution is Deep Deterministic Information Bottleneck (Yu
et al. 2021) that has been proved to reduce the superfluous
information during feature extraction. Before elaborating on
our solution, we introduce basic knowledge on DDIB below.
DDIB intends to adopt a neural network to parameterize
information bottleneck (Tishby, Pereira, and Bialek 1999),
which considers extracting information about a target signal
Ythrough a correlated observable X. The extracted infor-
mation is represented as a variable T. The information ex-
traction process can be formulated as:
maxLIB=I(T;Y)−βI(T;X). (2)
When the above objective is optimized with a neural net-
work,Tis the output of one hidden layer. To update theparameters of networks, the second item in Eq. (2) is cal-
culated with the differentiable matrix-based R ´enyi’sα-order
mutual information:
Iα(X;T) =Hα(X) +Hα(T)−Hα(X,T),(3)
whereHα(·)indicates the matrix-based analogue to R ´enyi’s
α-entropy and Hα(·,·)is the matrix-based analogue to
R´enyi’sα-order joint-entropy. More details of the matrix-
based R ´enyi’sα-order entropy functional can be found in
(Yu et al. 2019).
For the first item in Eq. (2), since I(T;Y) =H(Y)−
H(Y|T), the maximization of I(T;Y)is converted to
the minimization of H(Y|T). Given the training set
{xi,yi}N
i=1, the average cross-entropy loss is adopted to
minimize the H(Y|T):
1
NN/summationdisplay
i=1Et∼p(t|xi)[−logp(yi|t)], (4)
Therefore, DDIB indicates that the optimization of Infor-
mation Bottleneck in single modality can be achieved with
a cross-entropy loss and a differentiable mutual informa-
tion itemI(T;X). Obviously, the differentiable optimiza-
tion strategy of information bottleneck in DDIB can benefit
DSH methods in terms of superfluous information reduction.
Method
In this section, we first present the problem definition, and
then detail our DSIBH method. The optimization is finally
given. For illustration purposes, our DSIBH is applied in X-
ray images and radiology reports.
Notation and problem definition
Matrix and vector used in this paper are represented by bold-
face uppercase letter (e.g., G) and boldface lowercase let-
ter (e.g., g) respectively./bardbl·/bardbldenotes the 2-norm of vectors.
sign(·)is defined as the sign function, which outputs 1 if its
input is positive else outputs -1.
LetX1=/braceleftbig
x1
i/bracerightbigN
i=1andX2=/braceleftbig
x2
j/bracerightbigN
j=1symbolize X-
ray images and radiology reports in the training set, where
x1
i∈Rd1,x2
j∈Rd2. Their semantic labels that indicate
the existence of pathology are represented by Y={yl}N
l=1,
where yl={yl1,yl2,...,yld3}∈Rd3. Following (Cao et al.
2016; Jiang and Li 2017; Li et al. 2018), we define the se-
mantic affinities SN×Nbetween x1
iandx2
jusing semantic
labels. If x1
iandx2
jshare at least one category label, they
are semantically similar and Sij= 1. Otherwise, they are
semantically dissimilar and thus Sij= 0.
The goal of the proposed DSIBH is to learn hash functions
f1/parenleftbig
θ1;X1/parenrightbig
:Rd1→Rdcandf2/parenleftbig
θ2;X2/parenrightbig
:Rd2→Rdc,
which can map X-ray images and radiology reports as ap-
proximate binary codes G1andG2in the same continuous
space respectively. Later, binary codes can be generated by
applying a sign function to G1,2.
Meanwhile, hamming distance D/parenleftbig
g1
i,g2
j/parenrightbig
between hash
codes g1
iandg2
jneeds to indicate the semantic similaritySijbetween x1
iandx2
j, which can be formulated as:
Sij∝−D/parenleftbig
g1
i,g2
j/parenrightbig
. (5)
Information Bottleneck in Cross-modal Scenario
To improve the accuracy of CMR-based CAD, the super-
fluous information from the medical data in the hash code
learning procedure should be reduced via the information
bottleneck principle. Therefore, the information bottleneck
principle in single modality should be extended to the cross-
modal scenario, where one instance can own descriptions in
different modalities.
Analysis starts from the hash code learning processes for
X-ray images and radiology reports respectively. Follow-
ing the information bottleneck principle, the basic objective
functions can be formulated as:
maxLIB1=I/parenleftbig
G1;Y1/parenrightbig
−βI/parenleftbig
G1;X1/parenrightbig
,
maxLIB2=I/parenleftbig
G2;Y2/parenrightbig
−βI/parenleftbig
G2;X2/parenrightbig
. (6)
In cross-modal scenario, X-ray images and radiology re-
ports in the training set are collected to describe the com-
mon pathology. Therefore, the image-report pairs own the
same semantic label. To implement this idea, Eq. (6) is trans-
formed to:
maxLIB1=I/parenleftbig
G1;Y/parenrightbig
−βI/parenleftbig
G1;X1/parenrightbig
,
maxLIB2=I/parenleftbig
G2;Y/parenrightbig
−βI/parenleftbig
G2;X2/parenrightbig
, (7)
Furthermore, the same hash code should be assigned for the
paired samples to guarantee the consistency among different
modalities, which is achieved with the /lscript2loss:
minLCONS =E/bracketleftBig/vextenddouble/vextenddoubleg1−gy/vextenddouble/vextenddouble2/bracketrightBig
+E/bracketleftBig/vextenddouble/vextenddoubleg2−gy/vextenddouble/vextenddouble2/bracketrightBig
,(8)
whereGyrepresents the modality-invariant hash codes for
the image-report pairs.
Incorporating Eq. (7) and Eq. (8), the overall objective of
the information bottleneck principle in cross-modal scenario
is formulated as:
maxLIBC=/parenleftbig
I/parenleftbig
G1;Y/parenrightbig
+I/parenleftbig
G2;Y/parenrightbig/parenrightbig
(9)
−β/parenleftbig
I/parenleftbig
G1;X1/parenrightbig
+I/parenleftbig
G2;X2/parenrightbig/parenrightbig
−γ/parenleftBig
E/bracketleftBig/vextenddouble/vextenddoubleg1−gy/vextenddouble/vextenddouble2/bracketrightBig
+E/bracketleftBig/vextenddouble/vextenddoubleg2−gy/vextenddouble/vextenddouble2/bracketrightBig/parenrightBig
.
Deep Supervised Information Bottleneck Hashing
Following the information bottleneck principle in cross-
modal scenario (i.e., Eq. (9)), three variables including
Gy,G1andG2should be optimized. To obtain modality-
invariantGy, we build labNet fyto directly transform se-
mantic labels into the pair-level hash codes. The labNet is
formed by a two-layer Multi-Layer Perception (MLP) whose
nodes are 4096 and c. Then, we build imgNet f1and txtNet
f2as hash functions to generate hash codes G1andG2. For
X-ray images, we modify CNN-F (Chatfield et al. 2014)
to build imgNet with the consideration of network scale.
To obtaincbit length hash codes, the last fully-connectedlayer in the origin CNN-F is changed to a c-node fully-
connected layer. For radiology reports, we first use the multi-
scale network in (Li et al. 2018) to extract multi-scale fea-
tures and a two-layer MLP whose nodes are 4096 and cto
transform them into hash codes. Except the activation func-
tion of last layers is tanh to approximate the sign (·)func-
tion, other layers use ReLU as activation functions. To im-
prove generalization performance, Local Response Normal-
ization (LRN) (Krizhevsky, Sutskever, and Hinton 2012) is
applied between layers of all MLPs. One should note that the
application of CNN-F (Chatfield et al. 2014) and multi-scale
network (Li et al. 2018) is only for illustrative purposes; any
other networks can be integrated into our DSIBH as back-
bones of imgNet and txtNet.
As described before, semantic labels are encoded as hash
codesGy. To preserve semantic similarity, the loss function
of labNet is:
min
Gy,θyLy=Ly
1+ηLy
2 (10)
=−N/summationdisplay
l,j/parenleftbig
Slj∆lj−log/parenleftbig
1 +e∆lj/parenrightbig/parenrightbig
+ηN/summationdisplay
l=1/parenleftBig
/bardblgy
l−fy(θy;yl)/bardbl2/parenrightBig
,
s.t.Gy={gy
l}N
l=1∈{− 1,1}c
where∆lj=fy(θy;yl)Tfy/parenleftbig
θy;yj/parenrightbig
,fy(θy;yl)is the
output of labNet for yl,gy
lis the hash codes of fy(θy;yl)
handled bysign (·), andηaims to adjust the weight of loss
items.
The first term of Eq. (10) intends to minimize the nega-
tive log likelihood of semantic similarity with the likelihood
function, which is defined as follows:
p/parenleftbig
Slj|fy(θy;yl),fy/parenleftbig
θy;yj/parenrightbig/parenrightbig
=/braceleftbiggσ(∆lj) Slj= 1
1−σ(∆lj)Slj= 0,
(11)
whereσ(∆lj) =1
1+e−∆ljis the sigmoid function. Mean-
while, the second term restricts the outputs of labNet to ap-
proximate binary as the request of hash codes.
After the optimization of labNet, the modality-invariant
hash codeGyis obtained. The next step is to optimize the
imgNet and txtNet to generate G1andG2respectively fol-
lowing Eq. (9). For the first item in Eq. (9), DDIB interprets
it as a cross-entropy loss (i.e., Eq. (4)). In our implement, Gy
is also used as class-level weight in the cross-entropy loss,
which intends to make G1andG2inherent the semantic sim-
ilarity of the modality-invariant hash code. Specifically, the
non-redundant multi-label annotations are transformed into
Ny-class annotations{¯yl}Ny
l=1, and their corresponding hash
codes are regarded as the class-level weights. The weightedcross-entropy loss is formulated as:
min
θmLm
1=−1
NN/summationdisplay
iNy/summationdisplay
l¯yllog (ail), (12)
ail,exp/parenleftBig
(¯gy
l)Tgm
i/parenrightBig
/summationtextNy
l/primeexp/parenleftBig
(¯gy
l/prime)Tgm
i/parenrightBig,
wheremindicates the modality 1 or 2.
For the second item in Eq. (9), we adopt the differen-
tiable matrix-based R ´enyi’sα-order mutual information to
estimate:
min
θmLm
2=I(Gm;Xm). (13)
For the third item in Eq. (9), the /lscript2loss is directly used:
min
θmLm
3=N/summationdisplay
i=1/parenleftBig
/bardblgy
i−gm
i/bardbl2/parenrightBig
. (14)
By merging Eqs. (12), (13) and (14) together, we obtain
the loss function of imgNet (or txtNet), formulated as the
following minimization problem:
min
θmLm=Lm
1+βLm
2+γLm
3, (15)
whereβandγare hyper-parameters that are used to adjust
the weights of loss items.
Optimization
The optimization of our DSIBH includes two parts: learning
the modality-invariant hash code Gyand learning the hash
codes G1andG2for X-ray images and radiology reports
respectively. Learning Gyequals to optimize θy+. For hash
codes of modality m,θmneeds to be optimized. The whole
optimization procedure is summarized in Algorithm 1.
Forθyof labNet, Eq. (10) is derivable. Therefore, Back-
propagation algorithm (BP) with mini-batch stochastic gra-
dient descent (mini-batch SGD) method is applied to update
it. As for gy
l, we use Eq. (16) to update:
gy
l=sign (fy(θy;yl)). (16)
For imgNet and txtNet, we also use the BP with mini-
batch SGD method to update θ1andθ2.
Once Algorithm 1 converges, the well-trained imgNet and
txtNet with sign(·)are used to handle out-of-sample data
points from modality m:
gm
i=sign (fm(θm;xm
i)). (17)
Experiments
In this section, we first introduce the dataset used for assess-
ment and specify the experimental setting. Following this,
we demonstrate that the proposed DSIBH can achieve the
state-of-the-art performance on CMR-based CAD.Table 1: Comparison with baselines in terms of MAP on CMR-based CAD. The best results are marked with bold .
MethodX→R R →X
16 bits 32 bits 64 bits 128 bits 16 bits 32 bits 64 bits 128 bits
CCA (Hotelling 1992) 0.3468 0.3354 0.3273 0.3215 0.3483 0.3368 0.3288 0.3230
CMSSH (Bronstein et al. 2010) 0.4224 0.4020 0.3935 0.3896 0.3899 0.3967 0.3646 0.3643
SCM (Zhang and Li 2014) 0.4581 0.4648 0.4675 0.4684 0.4516 0.4574 0.4604 0.4611
STMH (Wang et al. 2015) 0.3623 0.3927 0.4211 0.4387 0.3980 0.4183 0.4392 0.4453
CMFH (Ding et al. 2016) 0.3649 0.3673 0.3736 0.3760 0.4130 0.4156 0.4303 0.4309
SePH (Lin et al. 2016) 0.4684 0.4776 0.4844 0.4903 0.4475 0.4555 0.4601 0.4658
DCMH (Jiang and Li 2017) 0.4834 0.4878 0.4885 0.4839 0.4366 0.4513 0.4561 0.4830
SSAH (Li et al. 2018) 0.4894 0.4999 0.4787 0.4624 0.4688 0.4806 0.4832 0.4833
EGDH (Shi et al. 2019) 0.4821 0.5010 0.4996 0.5096 0.4821 0.4943 0.4982 0.5041
DSIBH 0.5001 0.5018 0.5116 0.5172 0.4898 0.4994 0.4997 0.5084
Query X-ray images Retrieved  radiology reports via our DSIBH
Query radiology reports Retrieved X -ray images via our DSIBH(a)
(b)Theendotracheal tube tipis6cmabove
the carina .Nasogastric tube tipis
beyond theGEjunction andofftheedge
ofthefilm.Aleftcentral lineispresent in
thetipisinthemidSVC.Apacemaker is
noted ontheright inthelead projects
over the right ventricle .There is
probable scarring inboth lung apices .
There arenonew areas ofconsolidation .
There isupper zone redistribution and
cardiomegaly suggesting pulmonary
venous hypertension .There isno
pneumothorax .
There iscardiomegaly .Apacemaker is
present with thelead overlying theright
ventricle .Anapparent Swan -Ganz
catheter ispresent, with tipoverlying the
distal right pulmonary artery .There is
upper zone re-distribution, with mild
vascular plethora, butnoovert CHF.No
focal infiltrate isdetected .Possible trace
fluid attheleftcostophrenic angle .Ascompared totheprevious radiograph,
thepatient hasreceived aright -sided
chest tube .The tube isincorrect
position, the right lung isnow fully
expanded, there isnoevidence ofright
pneumothorax .Incomparison with thestudy of___,the
patient has taken abetter inspiration .
Continued enlargement ofthecardiac
silhouette with minimal central vascular
congestion .Right PICC lineisstable .No
evidence ofacute focal pneumonia .
Since prior radiograph from ___, the
mediastinal drain tube hasbeen removed .
There isnopneumothorax .Both lung
volumes arevery low.Bilateral, right side
more than left side, moderate
pulmonary edema has improved .
Widened cardiomediastinal silhouette is
more than itwas on___;however, this
appearance could beexacerbation from
lowlung volumes .Patient isstatus post
median sternotomy with intact sternal
sutures .Atelectasis; Pleural Effusion; 
PneumothoraxAtelectasis; Pleural Effusion; Support 
DevicesAtelectasis; CardiomegalyAtelectasis; Cardiomegaly; Pleural 
Effusion; Support DevicesAtelectasis; Cardiomegaly
Edema; Enlarged Cardiomediastinum ; 
Support DevicesEdema; Lung Opacity; Pleural Effusion; 
Pneumonia; Support DevicesCardiomegaly; Pleural Effusion; 
Support DevicesAtelectasis; Pleural Effusion; Support 
DevicesCardiomegaly; Edema; Pleural 
Effusion; Support Devices
Figure 1: The top 4 profiles retrieved by our DSIBH on the MIMIC-CXR dataset with 128 bits.
Experimental setting
The large-scale chest X-ray and radiology report dataset
MIMIC-CXR (Johnson et al. 2019) is used to evaluate the
performance of DSIBH. Some statistics of this dataset are
introduced as follows.
MIMIC-CXR1consists of chest X-ray images and radi-
ology reports sourced from the Beth Israel Deaconess Med-
ical Center between 2011-2016. Each radiology report is as-
sociated with at least one X-ray image and annotated with
a 14-dimensional label indicating the existence of pathol-
ogy or lack of pathology. To evaluate the performance of
CMR-based CAD, we adopt 73876 image-report pairs for
assessment. During the comparison process, radiology re-
ports are represented as bag-of-word vectors according to
the top 617 most-frequent words. In the testing phase, we
randomly sample 762 image-report pairs as query set and
regard the rest as retrieval set. In the training phase, 14000
pairs from the retrieval set are used as training set.
The proposed DSIBH is compared with nine state-of-
the-arts in hashing-based CMR including CCA (Hotelling
1992), CMSSH (Bronstein et al. 2010), SCM (Zhang and
Li 2014), STMH (Wang et al. 2015), CMFH (Ding et al.
1https://physionet.org/content/mimic-cxr/2.0.0/2016), SePH (Lin et al. 2016), DCMH (Jiang and Li 2017),
SSAH (Li et al. 2018), and EGDH (Shi et al. 2019). CCA,
STMH and CMFH are unsupervised approaches that depend
on data distributions, whereas the other six are supervised
methods that take semantic labels into account. For fair com-
parison with shallow-structure-based baselines, we use the
trainset of MIMIC-CXR to optimize a CNN-F network for
classification and extract 4096-dimensional features to rep-
resent X-ray images. We set η= 1,β= 0.1andγ= 1
for MIMIC-CXR as hyper-parameters. In the optimization
phase, the batch size is set as 128 and three Adam solvers
with different learning rates are applied (i.e., 10−3for lab-
Net,10−4.5for imgNet and 10−3.5for txtNet).
Mean average precision (MAP) is adopted to evaluate the
performance of hashing-based CMR methods. MAP is the
most widely used criteria metric to measure retrieval accu-
racy, which is computed as follows:
MAP =1
|Q||Q|/summationdisplay
i=11
rqiR/summationdisplay
j=1Pqi(j)δqi(j), (18)
where|Q|indicates the number of query set, rqirepresents
the number of correlated instances of query qiin database
set,Ris the retrieval radius, Pqi(j)denotes the precision ofAlgorithm 1: The Optimization Procedure of DSIBH
Input : X-ray images X1, radiology reports X2, semantic
labels Y, learning rates λy,λ1,λ2, and iteration numbers
Ty,T1,T2.
Output : Parameters θ1andθ2of imgNet and
txtNet.
1:Randomly initialize θy,θ1,θ2andGy.
2:repeat
3: foriter=1 toTydo
4: Updateθyby BP algorithm:
θy←θy−λy·∇θyLy
5: Update Gyby Eq. (16)
6: end for
7: foriter=1 toT1do
8: Updateθ1by BP algorithm:
θ1←θ1−λ1·∇θ1L1
9: end for
10: foriter=1 toT2do
11: Updateθ2by BP algorithm:
θ2←θ2−λ2·∇θ2L2
12: end for
13:until Convergence
the topjretrieved sample and δqi(j)indicates whether the
jthreturned sample is correlated with the ithquery entity. To
reflect the overall property of rankings, the size of database
set is used as the retrieval radius.
The efficacy of DSIBH in CMR-based CAD
CMR-based CAD stresses on two retrieval directions: using
X-ray images to retrieve radiology reports ( X→R) and
using radiology reports to retrieve X-ray images ( R→X).
In experiments, we set bit length as 16, 32, 64 and 128 bits.
Table 1 reports the MAP results on the MIMIC-CXR
dataset. As can be seen, unsupervised methods fail to pro-
vide reasonable retrieval results due to the neglect of se-
mantic information. CCA performs the worst among these
unsupervised methods due to the naive management of data
distribution. Compared with CCA, STMH and CMFH can
achieve a better retrieval accuracy, which we argue can
be attributed to the coverage of data correlation. By con-
trast, shallow-structure-based supervised methods includ-
ing CMSSH, SCM, and SePH achieve a large performance
gain over unsupervised methods by further considering se-
mantic information to express semantic similarity with hash
codes. Benefiting from the effect of nonlinear fitting abil-
ity and self-adjusting feature extraction ability, deep super-
vised methods including DCMH, SSAH and EGDH out-
perform the six shallow methods in the mass. Due to the
extra consideration of superfluous information reduction,
our DSIBH can achieve the best accuracy. Specifically,
compared with the recently proposed deep hashing method
EGDH by MAP, our DSIBH achieves average absolute in-
creases of 0.96%/0.47% on the MIMIC-CXR dataset.
Meanwhile, we also visualize the top 4 retrieved medical
profiles of our DSIBH on X→RandR→Xdirections
using the MIMIC-CXR dataset in Figure 1. These resultsconfirm our concern that DSIBH can retrieve pathology-
related heterogeneous medical data again.
Conclusion
In this paper, to preform computer-aided diagnosis (CAD)
based on the large-scale multi-modal medical data, the
cross-modal retrieval (CMR) technique based on semantic
hashing is introduced. Inspired by Deep Deterministic In-
formation Bottleneck, a novel method named Deep Super-
vised Information Bottleneck Hashing (DSIBH) is designed
to perform CMR-based CAD. Experiments are conducted
on the large-scale medical dataset MIMIC-CXR. Compared
with other state-of-the arts, our DSIBH can reduce the dis-
traction of superfluous information, which thus strengthens
the discriminability of hash codes in CMR-based CAD.
Acknowledgements
This work is partially supported by NSFC (62101179,
61772220), Key R&D Plan of Hubei
Province (2020BAB027) and Project of Hubei Univer-
sity School (202011903000002).
References
Bronstein, M. M.; Bronstein, A. M.; Michel, F.; and Para-
gios, N. 2010. Data fusion through cross-modality metric
learning using similarity-sensitive hashing. In 2010 IEEE
computer society conference on computer vision and pattern
recognition , 3594–3601. IEEE.
Cao, Y .; Long, M.; Wang, J.; and Liu, S. 2017. Collec-
tive deep quantization for efficient cross-modal retrieval. In
Thirty-First AAAI Conference on Artificial Intelligence .
Cao, Y .; Long, M.; Wang, J.; and Zhu, H. 2016. Correlation
autoencoder hashing for supervised cross-modal search. In
Proceedings of the 2016 ACM on International Conference
on Multimedia Retrieval , 197–204. ACM.
Chatfield, K.; Simonyan, K.; Vedaldi, A.; and Zisserman, A.
2014. Return of the devil in the details: Delving deep into
convolutional nets. arXiv preprint arXiv:1405.3531 .
de La Torre, J.; Valls, A.; and Puig, D. 2020. A deep learn-
ing interpretable classifier for diabetic retinopathy disease
grading. Neurocomputing , 396: 465–476.
Ding, G.; Guo, Y .; Zhou, J.; and Gao, Y . 2016. Large-
scale cross-modality search via collective matrix factoriza-
tion hashing. IEEE Transactions on Image Processing ,
25(11): 5427–5440.
Erfankhah, H.; Yazdi, M.; Babaie, M.; and Tizhoosh, H. R.
2019. Heterogeneity-aware local binary patterns for retrieval
of histopathology images. IEEE Access , 7: 18354–18367.
Fang, J.; Fu, H.; and Liu, J. 2021. Deep triplet hashing net-
work for case-based medical image retrieval. Medical Image
Analysis , 69: 101981.
Hotelling, H. 1992. Relations between two sets of variates.
InBreakthroughs in statistics , 162–190. Springer.
Hu, H.; Xie, L.; Hong, R.; and Tian, Q. 2020. Creating
Something From Nothing: Unsupervised Knowledge Distil-
lation for Cross-Modal Hashing. In 2020 IEEE/CVF Confer-
ence on Computer Vision and Pattern Recognition (CVPR) .Hu, Z.; Liu, X.; Wang, X.; Cheung, Y .-m.; Wang, N.; and
Chen, Y . 2019. Triplet Fusion Network Hashing for Un-
paired Cross-Modal Retrieval. In Proceedings of the 2019
on International Conference on Multimedia Retrieval , 141–
149.
In´es, A.; Dom ´ınguez, C.; Heras, J.; Mata, E.; and Pascual, V .
2021. Biomedical image classification made easier thanks to
transfer and semi-supervised learning. Computer Methods
and Programs in Biomedicine , 198: 105782.
Jiang, Q.-Y .; and Li, W.-J. 2017. Deep cross-modal hashing.
InProceedings of the IEEE conference on computer vision
and pattern recognition , 3232–3240.
Johnson, A. E.; Pollard, T. J.; Greenbaum, N. R.; Lungren,
M. P.; Deng, C.-y.; Peng, Y .; Lu, Z.; Mark, R. G.; Berkowitz,
S. J.; and Horng, S. 2019. MIMIC-CXR-JPG, a large pub-
licly available database of labeled chest radiographs. arXiv
preprint arXiv:1901.07042 .
Krizhevsky, A.; Sutskever, I.; and Hinton, G. E. 2012. Im-
agenet classification with deep convolutional neural net-
works. In Advances in neural information processing sys-
tems, 1097–1105.
Li, C.; Deng, C.; Li, N.; Liu, W.; Gao, X.; and Tao, D.
2018. Self-supervised adversarial hashing networks for
cross-modal retrieval. In Proceedings of the IEEE con-
ference on computer vision and pattern recognition , 4242–
4251.
Lin, Z.; Ding, G.; Han, J.; and Wang, J. 2016. Cross-view re-
trieval via probability-based semantics-preserving hashing.
IEEE transactions on cybernetics , 47(12): 4342–4355.
Liu, S.; Qian, S.; Guan, Y .; Zhan, J.; and Ying, L.
2020. Joint-modal Distribution-based Similarity Hashing
for Large-scale Unsupervised Deep Cross-modal Retrieval.
InProceedings of the International ACM SIGIR Conference
on Research and Development in Information Retrieval ,
1379–1388.
Shen, F.; Gao, X.; Liu, L.; Yang, Y .; and Shen, H. T. 2017.
Deep asymmetric pairwise hashing. In Proceedings of the
ACM international conference on Multimedia , 1522–1530.
Shi, X.; Su, H.; Xing, F.; Liang, Y .; Qu, G.; and
Yang, L. 2020. Graph temporal ensembling based semi-
supervised convolutional neural network with noisy labels
for histopathology image analysis. Medical Image Analysis ,
60: 101624.
Shi, Y .; You, X.; Zheng, F.; Wang, S.; and Peng, Q. 2019.
Equally-guided discriminative hashing for cross-modal re-
trieval. In Proceedings of the 28th International Joint Con-
ference on Artificial Intelligence , 4767–4773.
Song, G.; Tan, X.; Zhao, J.; and Yang, M. 2021. Deep Ro-
bust Multilevel Semantic Hashing for Multi-Label Cross-
Modal Retrieval. Pattern Recognition , 108084.
Tishby, N.; Pereira, F. C.; and Bialek, W. 1999. The infor-
mation bottleneck method. 368–377.
Wang, D.; Gao, X.; Wang, X.; and He, L. 2015. Seman-
tic topic multimodal hashing for cross-media retrieval. In
Twenty-Fourth International Joint Conference on Artificial
Intelligence .Wang, J.; Zhang, T.; Sebe, N.; Shen, H. T.; et al. 2017. A
survey on learning to hash. IEEE transactions on pattern
analysis and machine intelligence , 40(4): 769–790.
Wang, K.; Yin, Q.; Wang, W.; Wu, S.; and Wang, L. 2016.
A comprehensive survey on cross-modal retrieval. arXiv
preprint arXiv:1607.06215 .
Wang, L.; Zhu, L.; Yu, E.; Sun, J.; and Zhang, H. 2019.
Fusion-supervised deep cross-modal hashing. In IEEE Inter-
national Conference on Multimedia and Expo , 37–42. IEEE.
Xie, D.; Deng, C.; Li, C.; Liu, X.; and Tao, D. 2020.
Multi-Task Consistency-Preserving Adversarial Hashing for
Cross-Modal Retrieval. IEEE Transactions on Image Pro-
cessing , 29: 3626–3637.
Yang, E.; Yao, D.; Cao, B.; Guan, H.; Yap, P.-T.; Shen, D.;
and Liu, M. 2020. Deep disentangled hashing with momen-
tum triplets for neuroimage search. In International Confer-
ence on Medical Image Computing and Computer-Assisted
Intervention , 191–201. Springer.
Yao, H.-L.; Zhan, Y .-W.; Chen, Z.-D.; Luo, X.; and Xu,
X.-S. 2021. TEACH: Attention-Aware Deep Cross-Modal
Hashing. In Proceedings of the 2021 International Confer-
ence on Multimedia Retrieval , ICMR ’21, 376–384. New
York, NY , USA: Association for Computing Machinery.
ISBN 9781450384636.
Yu, J.; Zhou, H.; Zhan, Y .; and Tao, D. 2021. Deep Graph-
neighbor Coherence Preserving Network for Unsupervised
Cross-modal Hashing. In Proceedings of the AAAI Confer-
ence on Artificial Intelligence , volume 35, 4626–4634.
Yu, S.; Giraldo, L. G. S.; Jenssen, R.; and Principe, J. C.
2019. Multivariate Extension of Matrix-Based R ´enyi’s\
α-Order Entropy Functional. IEEE transactions on pattern
analysis and machine intelligence , 42(11): 2960–2966.
Yu, X.; Yu, S.; and Pr ´ıncipe, J. C. 2021. Deep Deterministic
Information Bottleneck with Matrix-Based Entropy Func-
tional. In ICASSP 2021-2021 IEEE International Confer-
ence on Acoustics, Speech and Signal Processing (ICASSP) ,
3160–3164. IEEE.
Zhan, Y .-W.; Luo, X.; Wang, Y .; and Xu, X.-S. 2020. Su-
pervised Hierarchical Deep Hashing for Cross-Modal Re-
trieval. In Proceedings of the 28th ACM International Con-
ference on Multimedia , MM ’20, 3386–3394. New York,
NY , USA: Association for Computing Machinery. ISBN
9781450379885.
Zhang, D.; and Li, W.-J. 2014. Large-scale supervised mul-
timodal hashing with semantic correlation maximization. In
Twenty-Eighth AAAI Conference on Artificial Intelligence .
Zhang, J.; Xie, Y .; Xia, Y .; and Shen, C. 2019. Attention
residual learning for skin lesion classification. IEEE trans-
actions on medical imaging , 38(9): 2092–2103.
Zhen, L.; Hu, P.; Wang, X.; and Peng, D. 2020. Deep Super-
vised Cross-Modal Retrieval. In 2019 IEEE/CVF Confer-
ence on Computer Vision and Pattern Recognition (CVPR) .
Zhu, L.; Lu, X.; Cheng, Z.; Li, J.; and Zhang, H. 2020. Flex-
ible multi-modal hashing for scalable multimedia retrieval.
ACM Transactions on Intelligent Systems and Technology
(TIST) , 11(2): 1–20.