Classiﬁcation of Histopathological Images of Penile Cancer using

DenseNet and Transfer Learning

Marcos Gabriel Mendes Lauande

, Amanda Mara Teles

, Leandro Lima da Silva

Caio Eduardo Falc

ao Matos

, Geraldo Braz J

unior

, Anselmo Cardoso de Paiva

ao Dallyson Sousa de Almeida

, Rui Miguel Gil da Costa Oliveira

2,3

, Haissa Oliveira Brito

Ana Gis

elia Nascimento

, Ana Clea Feitosa Pestana

2,4

, Ana Paula Silva Azevedo dos Santos

and Fernanda Ferreira Lopes

Computer Applied Group (NCA), Federal University of Maranh

ao (UFMA), S

ao Lu

ıs - MA, Brazil

Graduate Program in Adult Health - PPGSAD, Federal University of Maranh

ao (UFMA) S

ao Lu

ıs - MA, Brazil

Centre for Research and Technology of Agro-Environmental and Biological Sciences (CITAB), Inov4Agro,

University of Tr

as-os-Montes and Alto Douro (UTAD), Vila Real, Portugal

Department of Pathology, Presidente Dutra University Hospital - Federal University of Maranh

ao (UFMA),

ao Lu

ıs - MA, Brazil

marcos.lauande, geraldo.braz, anselmo.paiva, joao.dallyson, rui.costa, haissa.brito, ana.giselia, ana.pestana, ana.azevedo,

Keywords:

Histopathology, Penile Cancer, Deep Learning, Deep Features, Convolutional Neural Network, Transfer

Learning, Data Augmentation, Contrast Limited Adaptative Histogram Equalization.

Abstract:

Penile cancer is a rare tumor that accounts for 2% of cancer cases in men in Brazil. Histopathological analyzes

are commonly used in its diagnosis, making it possible to assess the degree of the disease, its evolution, and its

nature. About a decade ago, scientiﬁc works in the ﬁeld of deep learning were developed to help pathologists

make decisions quickly and reliably, opening up possibilities for new contributions to improve such a complex

and time-consuming activity for these professionals. In this work, we present the development of a method

that uses a DenseNet to diagnose penile cancer in histopathological images, and the construction of a dataset

(via the Legal Amazon Penis Cancer Project) used to validate this method. In the experiments performed, an

F1-Score of up to 97.39% and a sensitivity of up to 98.33% were achieved in this binary classiﬁcation problem

(normal or squamous cell carcinoma).

1 INTRODUCTION

Cancer is related to disordered cell growth. Depend-

ing on the degree of malignancy, it can invade adja-

cent tissues or organs, leading the patient to death if

there is no adequate early treatment. Penile cancer

is a rare tumor and represents 0.4% to 0.6% of all

cancers in Europe and North America but is consid-

erably more common in developing countries in Latin

America, Africa, and Asia (Douglawi and Masterson,

2017). According (INCA, 2021), this disease has a

higher incidence in men aged 50 years and over in

Brazil, being more common in the North and North-

east regions of the country and corresponding to 2%

of cancer cases in men. A report from the state of

Maranh

ao, Brazil, indicates an age-standardized inci-

dence rate of 6.15 per 100,000 (Coelho et al., 2018),

which is very worrying. This type of tumor is linked

to some factors, such as lack of hygiene, human papil-

lomavirus (HPV) infection, the presence of phimosis,

and risky sexual behavior (Vieira et al., 2020).

According to (ACS, 2021), and (Thomas et al.,

2021), one of the exams that can be indicated for the

diagnosis of this disease is the histopathological anal-

ysis of tissues collected through biopsy, which con-

sists of the microscopic evaluation of very ﬁne tis-

sues extracted from the region of interest. Before be-

ing taken for analysis, these tissues are stained using

Eosin and Hematoxylin, then placed on glass slides

(Neto, 2012). The pathologist veriﬁes the structure

of tissue cells understanding the evolution, subtype,

and extent of the disease, making it possible to make

safer decisions about the type of treatment or surgery

to be prescribed. However, according to (Melo et al.,

2020), this activity tends to be very complex and time-

976

Lauande, M., Teles, A., Lima da Silva, L., Matos, C., Braz Júnior, G., Cardoso de Paiva, A., Sousa de Almeida, J., Oliveira, R., Brito, H., Nascimento, A., Pestana, A., Santos, A. and Lopes, F.

Classiﬁcation of Histopathological Images of Penile Cancer using DenseNet and Transfer Learning.

DOI: 10.5220/0010893500003124

In Proceedings of the 17th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2022) - Volume 4: VISAPP, pages

976-983

ISBN: 978-989-758-555-5; ISSN: 2184-4321

consuming because the process is detailed and has

a subjective conclusion, depending a lot on the pro-

fessional’s expertise. Because of this, software-based

solutions can help these professionals, bringing more

reliability, enabling a faster diagnosis, and the sooner

the disease is detected and together with adequate

treatment, the greater the probability of cure.

With the evolution of machine learning techniques

and image processing applied to medical images, sev-

eral related scientiﬁc research based on diagnosis

using histopathological images has been developed

(Srinidhi et al., 2021), such as for breast cancer (Cruz-

Roa et al., 2014) , prostate (Linkon et al., 2021), liver

(Kiani et al., 2020), colon (Sarwinda et al., 2021) and

lung (Wei et al., 2019). Thus, deep learning archi-

tectures, speciﬁcally convolutional neural networks

(CNNs), ended up becoming quite popular for this

type of problem. This popularity was due to its perfor-

mance, in certain cases, superior to traditional meth-

ods of feature extraction and classiﬁers not based on

neural networks; this can be seen in the results ob-

tained in (Spanhol et al., 2016a), for example. How-

ever, concerning cancer diagnosis based on images

at cellular levels of the penile tissue, the lack of a

dataset that can support the researchers’ experiments

becomes an impediment to the development of spe-

cialized methods.

In order to ﬁll this gap, this article proposes an au-

tomated penile cancer diagnosis method that analyzes

the histopathological images and determines whether

there is cancer or not. This method is based on

transfer learning that uses a pre-trained DenseNet-201

type convolutional network for the ImageNet chal-

lenge with an image preprocessing step using an algo-

rithm known as CLAHE (Contrast Limited Adaptive

Histogram Equalization). In addition, the developed

method was validated in a dataset built through the

Legal Amazon Penis Cancer Project, comprising 194

histopathological images classiﬁed as normal or can-

cer (squamous cell carcinoma).

Our contributions are listed below: (a) the propo-

sition of a new method for automated diagnosis of pe-

nile cancer based on deep learning using microscope

images; (b) the construction of a new dataset of penile

cancer histopathological images that served to vali-

date the developed method; and (c) analysis and com-

parison of the proposed method with another one built

by the same author based on deep features and tradi-

tional classiﬁers.

This paper’s remainder is organized as follows:

Section II presents the related works that served as

the basis for the research that resulted in this paper,

Section III explains how the histopathological im-

ages of penile cancer were acquired and details about

the proposed method based on deep learning, Sec-

tion IV, which presents an analysis of the results ob-

tained through the method and other experiments car-

ried out; and, ﬁnally, Section V, which concludes the

work with some ﬁnal considerations.

2 RELATED WORKS

Computer recognition of medical images, speciﬁcally

cancer histopathological images, is a widely explored

research topic, with many works available to obtain

theoretical foundations that served as the basis for

the entire development of a methodology and experi-

ments.

In (Filipczuk et al., 2013), the Hough Transform

was used to detect cell nuclei in histopathological

images, and these had their characteristics extracted

and used as input for an SVM (Support Vector Ma-

chine) classiﬁer. In work described in (Spanhol et al.,

2016b), feature extraction techniques such as Lo-

cal Binary Patterns, Gray-Level Co-Occurrence Ma-

trices, and Local Phase Quantization were used in

several experiments with traditional classiﬁers in a

dataset consisting of 7,909 breast cancer images.

Over time, traditional feature extraction tech-

niques were replaced by techniques based on deep

learning for this type of problem, with better results.

In (Sharma et al., 2017), a CNN architecture was

constructed and applied for cancer classiﬁcation and

necrosis detection using a gastric cancer histopatho-

logical image dataset. This work showed superior

performance compared to the Random Forest classi-

ﬁer.

The use of preprocessing techniques is essential

to improve image classiﬁcation task performance. A

work developed by (Sarwinda et al., 2021) demon-

strated the use of the CLAHE algorithm together with

the ResNet-18 and ResNet-50 neural networks for a

binary classiﬁcation problem (malignant or benign).

In this case, a database of large intestine tissue im-

ages had its samples converted to grayscale and pre-

processed using the CLAHE algorithm, substantially

improving the classiﬁcation results.

The transfer of learning technique in the classiﬁ-

cation of histopathological images has become very

promising and popular, especially in cases where ex-

periments are restricted to less powerful computers

and datasets with few samples. This procedure con-

sists of using pre-trained networks in a given do-

main in a similar or different one. The application of

this technique can be seen in (Spanhol et al., 2017),

(Boumaraf et al., 2021) and (Choudhary et al., 2021),

and in this last work cited, the pre-trained network

Classiﬁcation of Histopathological Images of Penile Cancer using DenseNet and Transfer Learning

977

had its less important weights removed, improving the

overall result of the model. In addition, data augmen-

tation can be used to circumvent the sample quan-

tity limitation, the class imbalance, and the overﬁt-

ting problem, which can considerably affect the per-

formance of deep learning techniques, as they per-

form better on image bases with thousands or mil-

lions of samples, such as ImageNet (Russakovsky

et al., 2015). This technique causes the images to

undergo some kind of transformation, thus produc-

ing new samples, works such as those reported in

(Rakhlin et al., 2018) and (Tellez et al., 2019) showed

very promising results when using it.

Finally, these works are a sample of the growing

diversity of software-based studies with the purpose

of automating the activity of analyzing histopatholog-

ical images that represent state of the art in this ﬁeld

of research, being the foundation for the application

of a specialized diagnostic method in a type of cancer

image not explored in the literature.

3 MATERIALS AND METHOD

This section details the construction process of the

dataset of penile cancer histopathological images

used and the proposed method (Figure 1) that has

its following steps: image acquisition; image prepro-

cessing using the CLAHE (Contrast Limited Adaptive

Histogram Equalization) algorithm; transfer learn-

ing and classiﬁcation using an ImageNet pre-trained

model based on the DenseNet convolutional network

architecture; evaluation of the results on the subset of

test image.

3.1 Histopathological Images Dataset

Construction

The dataset of images provided by the Legal Ama-

zon Penis Cancer Project, which served to validate the

proposed method, consists of 194 RGB images with

a resolution of 2048x1536 pixels. These ﬁles were

grouped by magniﬁcation and pathological classiﬁca-

tion according to Table 1.

Table 1: Distribution of images according to magniﬁcation

and pathological classiﬁcation.

Category/Magniﬁcation 40X 100X

Normal 40 40

Cancer (squamous cell carcinoma) 57 57

Total 97 97

The image capture process was carried out in

2021 using penile tissue samples representing tumors

and adjacent non-tumor areas, stained with hema-

toxylin and eosin stored at the Maranh

ao Tumour and

DNA Biobank. Two graduate students photographed

the samples using a high-deﬁnition camera (Leica

ICC50 HD) coupled to a brightﬁeld microscope (Le-

ica DM500). With the aid of speciﬁc software (Leica

Aperio ImageScope), the images were analyzed and

classiﬁed by two pathologists as penile cancer or non-

tumor tissue, according to the international classiﬁca-

tion of penile tumors (Epstein et al., 2020). Some

examples of these images are shown in Figure 2.

3.2 Preprocessing

Before training or testing the model based on convo-

lutional neural networks, we veriﬁed that the images

had differences in light distribution. To minimize the

effect, we propose a preprocessing via CLAHE (Con-

trast Limited Adaptive Histogram Equalization) algo-

rithm presented in (Zuiderveld, 1994).

According to (Kumar and Shaik, 2015) CLAHE

is an evolution of Adaptive Histogram Equalization

(AHE) and its most basic precursor, Histogram Equal-

ization (HE). Created for medical images, this algo-

rithm is easy to implement and has good results when

applied to microscopic images that are most often af-

fected by low lighting effects at the time of acquisition

(Kumar and Shaik, 2015). The main difference com-

pared to its predecessors is in the way the histogram

equalization is applied, not being on the whole image

at once but on regions of adjustable size. The image

is divided into blocks that are adaptively equalized,

then combined and treated to avoid edge effects using

bilinear interpolation.

In this work, initially, the images are resized to

224x224 pixels and converted from RGB to YUV.

This transformation is necessary because all channels

in RGB space carry color information. Therefore, un-

desirable effects on image colors would be noticeable

after the equalization of these channels. Thus, the

CLAHE algorithm was applied to the Y channel of

the YUV color space (Vill

an, 2019). This channel

represents the intensity information that needs to be

improved. After this operation, the equalized images

were converted to RGB. An example of this technique

can be seen in Figure 3, where it is possible to see how

evident the cellular structures of the tissue are after

the operation.

3.3 Transfer Learning with

Convolutional Networks

Convolutional Neural Networks are special architec-

tures inspired by the biological mechanisms of vi-

VISAPP 2022 - 17th International Conference on Computer Vision Theory and Applications

978

Figure 1: Proposed method.

Figure 2: Examples of histopathological images of penile cancer by category and magniﬁcation.

Figure 3: This example demonstrates the application of the

CLAHE enhancement algorithm on a histopathological im-

age of penile cancer.

sion of living beings, being used in classiﬁcation,

segmentation and object recognition tasks, for exam-

ple. What makes them different from other neural

network-based architectures are convolution opera-

tions in some of their layers. Still, in a complementary

way, their structure is usually composed of other lay-

ers, such as pooling and fully connected layers. The

ﬁrst publication of this type of neural network was

made by (Lecun et al., 1998). Still, they only became

popular through the work of (Krizhevsky et al., 2012)

for the 2012 ImageNet challenge as commented on in

(Zhang et al., 2021), after that, many other new archi-

tectures emerged, being widely applied in the context

of medical image analysis.

The complete training of a CNN started with ran-

dom weights depends on an extremely large amount

of images and many computational resources to pro-

cess them. The quantity of samples tends to be a

problem in the case of medical images, as there is

an entire technical procedure to acquire this data in a

correct and controlled manner to produce quality arti-

facts. Furthermore, there is another factor which is the

availability of patients in a particular medical condi-

tion (Morid et al., 2021). To get around this problem,

the use of transfer learning technique is the most indi-

cated, in this approach the weights learned by a neu-

ral network in a given problem domain can be reused

in another different domain, but as the classiﬁcation

layer is speciﬁc for the original problem, it is neces-

sary to change the last layer to adapt this CNN archi-

tecture to a new problem. According to (Tajbakhsh

et al., 2016), the initial layers of convolutional net-

works learn low-level characteristics that are gener-

ally applicable to most computer vision tasks. In con-

trast, those from deeper layers learn high-level char-

acteristics speciﬁc to the problem domain by which

they are being applied. Based on these statements, an

auxiliary technique can be used together with trans-

fer learning to improve training results, called ﬁne-

tuning, which consists of training the aggregated lay-

ers and those from the pre-trained CNN.

DenseNets were presented in (Huang et al., 2017),

their main objectives are: to reduce the so-called van-

ishing gradient effect that often occurs in deep neural

networks, in addition to strengthening the propaga-

tion of features and signiﬁcantly reducing the number

of parameters to be learned. The main characteris-

tic of this type of CNN is how data is sent between

the layers, each of which has connections with the

other layers later, something very similar to ResNets.

A DenseNet network is composed of several dense

Classiﬁcation of Histopathological Images of Penile Cancer using DenseNet and Transfer Learning

979

blocks that are separated by a transition layer that

aims to reduce the size of the generated feature maps

that will be sent to the next layers. There are several

versions of DenseNets from the original work (Huang

et al., 2017), numbered by the number of layers; in

this case, the architecture used in this work has ex-

actly 201 layers by default.

The proposed method uses a pre-trained model

through the DenseNet-201 architecture in the Ima-

geNet image base available in the Keras library writ-

ten in Python. The use of this CNN type is jus-

tiﬁed because it had a superior performance in ex-

periments performed concerning other architectures

such as Xception and InceptionResnetV2, for exam-

ple. This model was then trained using an Nvidia

Geforce 3060 RTX in the penile cancer histopatholog-

ical image dataset for both types of magniﬁcation in

accordance with the transfer-learning technique. The

classiﬁcation layers were removed. The other layers

of the network remained frozen. Three additional lay-

ers were added: an average pooling layer, a dense

layer that uses the Relu activation function with 256

neurons, and an output layer that uses the Softmax al-

gorithm. Furthermore, a dropout with an empirically

deﬁned probability of 0.35 has been added between

the last two layers. Figure 4 shows the entire CNN ar-

chitecture with the dense blocks, its connections be-

tween the transition layers, and the layers added by

the author that are suitable for this binary classiﬁca-

tion problem.

Each training was performed in 35 epochs using

the Adam optimizer conﬁgured with a learning rate

of 0.0001. The batch size was adjusted to 32, and

some experiments used data augmentation (horizontal

and vertical ﬂip and random rotation up to 90º). Af-

ter each training, the resulting model was ﬁne-tuned

to improve the results. All layers of this network

were unfrozen, and the model was retrained with eight

epochs using a learning rate equal to 0.00001.

4 RESULTS AND DISCUSSION

After preprocessing the color images, the model was

trained using k-fold cross-validation to assess the gen-

eralizability of the model. Each experiment was con-

ﬁgured for ﬁve folds in a stratiﬁed way, and for each

of the ﬁve rounds, a different fold of images is se-

lected for testing; the others are used to compose the

training and validation partitions, respectively 80%

and 20% also in a stratiﬁed way. Table 2 presents

the experiments performed that demonstrate the in-

ﬂuence of preprocessing and data augmentation on

the results. In order to evaluate the proposed method,

the following metrics were used: accuracy, sensitiv-

ity, speciﬁcity, precision, and F1-Score, with the latter

indicator being the criterion used to indicate the best

model.

Table 2: List of experiments performed to demonstrate the

contribution of preprocessing with CLAHE and data aug-

mentation to method results.

Experiments

No. Mag Preprocessing Augmentation

1 40X Raw No

2 100X Raw No

3 40X CLAHE No

4 100X CLAHE No

5 40X Raw Yes

6 100X Raw Yes

7 40X CLAHE Yes

8 100X CLAHE Yes

As shown in Table 3, the application of the

CLAHE contributed to the method obtaining a good

result for the 40X magniﬁcation based on the F1-

Score metric, which is the harmonic mean between

precision and recall; therefore, for experiment 3, the

result was 97.39%(+/-2.13). For the 100X magniﬁ-

cation, it is veriﬁed that the data augmentation con-

tributed together with the CLAHE algorithm in exper-

iment 8, which obtained 97.31 (+/-3.62) of F1-Score.

In addition to good F1-Scores, these trained mod-

els had signiﬁcant results regarding recall. This met-

ric reports the proportion of images that have cancer

and that were rated positively. Therefore, experiment

3, carried out on images of 40X magniﬁcation, had

the result 98.33%(+/-3.33), being more stable with

a smaller standard deviation than other experiments;

for the 100X magniﬁcation, experiment 8 resulted in

98.18%(+/-3.64).

In addition to the experiments carried out to ver-

ify the performance of the method proposed in this

dataset of histopathological images, the author exper-

imented with another similar method based on the

deep features technique. DenseNet was used to ex-

tract features from the resized images (224x244 pix-

els) that were highlighted by the CLAHE algorithm.

Then the resulting feature vectors served as input

to a classiﬁer in the training and testing steps. To

ﬁnd a good model among several classiﬁers (Deci-

sion Tree, Random Forest, and K-Nearest Neighbors),

we used the GridSearch algorithm with stratiﬁed 5x5

folds nested cross-validation. The results by magniﬁ-

cation and classiﬁer can be seen in table 4. The re-

sults, in this case, were very promising, especially

for the selected model trained in the KNN classi-

ﬁer for 40X magniﬁcation, which had its indicators

slightly higher than the others, obtaining an F1-Score

VISAPP 2022 - 17th International Conference on Computer Vision Theory and Applications

980

Figure 4: DenseNet-201 architecture along with the layers added by the author: shortcut connections are pertinent to each

dense block, and the operations between blocks are the transition layers; ﬁnally, additional layers were included in order to

improve and adapt the network to the binary classiﬁcation problem.

Table 3: Results of the proposed experiments.

Experiment No. Accuracy(%) Recall(%) Especiﬁcity(%) Precision(%) F1-Score(%)

1 92.74(+/-4.26) 96.52(+/-4.27) 87.50(+/-7.91) 91.79(+/-5.28) 93.98(+/-3.52)

2 92.89(+/-5.06) 100.00(+/-0.00) 82.50(+/-12.75) 89.81(+/-6.69) 94.50(+/-3.72)

3 96.89(+/-2.54) 98.33(+/-3.33) 95.00(+/-6.12) 96.67(+/-4.08) 97.39(+/-2.13)

4 94.89(+/-3.16) 100.00(+/-0.00) 87.50(+/-7.91) 92.27(+/-4.55) 95.92(+/-2.44)

5 91.68(+/-6.37) 92.88(+/-6.77) 90.00(+/-9.35) 93.16(+/-6.71) 92.87(+/-5.53)

6 93.89(+/-3.72) 98.33(+/-3.33) 87.50(+/-7.91) 92.14(+/-4.56) 95.06(+/-2.97)

7 92.74(+/-4.26) 94.55(+/-7.27) 90.00(+/-5.00) 93.26(+/-3.48) 93.72(+/-4.01)

8 96.84(+/-4.21) 98.18(+/-3.64) 95.00(+/-6.12) 96.52(+/-4.27) 97.31(+/-3.62)

of 95.47%(+/-5.23). For the 100X magniﬁcation, the

model trained by the RF classiﬁer obtained an F1-

Score of 95.71%(+/-2.89). Despite these results, the

proposed method did better according to the compar-

ison presented in table 5.

Complementary, some experiments were carried

out on the test images to verify which regions the

convolutional neural network took into account to

make a given classiﬁcation decision. This veriﬁ-

cation was performed based on the results of the

Gradient-weighted Class Activation Mapping algo-

rithm, known as Grad-CAM (Selvaraju et al., 2019),

which is based on the gradient data of the last convo-

lutional layer of a CNN. This approach allowed a heat

map type image superimposed on each image clas-

siﬁed by the DenseNet network. Examples of these

experiments can be seen in Figure 5 for the 40X mag-

niﬁcation images and in Figure 6 for the 100X mag-

niﬁcation images. In this case, we conclude that the

neural network considered the tissue edge regions in

all cases as the most important for classiﬁcation.

5 CONCLUSION

In this work, a specialized method based on deep

learning and image enhancement for the problem of

binary classiﬁcation of penile cancer histopatholog-

Figure 5: Grad-CAM algorithm result when overlaying the

heat maps with the evaluated 40X magniﬁcation images.

ical images was presented. As discussed above, it

achieved promising results with an F1-Score of up

to 97.39%(+/-2.13) on an imaging database used for

the ﬁrst time in experiments to automate image-based

medical diagnoses.

As future work, we suggest to use techniques

that can better take advantage of the characteristics

of high-resolution images, such as patch extraction

methods and the use of weakly supervised learning

through multiple instance learning techniques, for ex-

ample. We pretend, furthermore, to update the dataset

Classiﬁcation of Histopathological Images of Penile Cancer using DenseNet and Transfer Learning

981

Table 4: Additional results for ablation experiments based on deep features and on the use of GridSearch for the selection of

models. The classiﬁers were represented by the following abbreviations: DT = Decision Tree; RF = Randon Forest; and KNN

= K-Nearest Neighbors.

Mag. Classiﬁer Accuracy(%) Recall(%) Especiﬁcity(%) Precision(%) F1-Score(%)

40X DT 87.68(+/-3.99) 91.36(+/-9.14) 82.50(+/-15.00) 89.59(+/-8.63) 89.69(+/-3.42)

40X RF 93.84(+/-1.92) 98.33(+/-3.33) 87.50(+/-7.91) 92.14(+/-4.56) 94.98(+/-1.34)

40X KNN 94.84(+/-5.77) 94.70(+/-7.20) 95.00(+/-6.12) 96.46(+/-4.39) 95.47(+/-5.23)

100X DT 78.47(+/-5.48) 82.73(+/-8.87) 72.50(+/-9.35) 81.44(+/-4.89) 81.75(+/-5.10)

100X RF 94.84(+/-3.33) 98.18(+/-3.64) 90.00(+/-5.00) 93.44(+/-3.32) 95.71(+/-2.89)

100X KNN 91.84(+/-8.27) 94.70(+/-7.20) 87.50(+/-13.69) 92.18(+/-7.90) 93.27(+/-6.78)

Table 5: Comparison between the proposed method (line 1

and 2) and another one created by the same author based on

deep features (line 3 and 4).

Mag. Feat. Ext. Classiﬁer F1-Score(%)

40X DenseNet DenseNet 97.39(+/-2.13)

100X DenseNet DenseNet 97.31(+/-3.62)

40X DenseNet KNN 95.47(+/-5.23)

100X DenseNet RF 95.71(+/-2.89)

Figure 6: Grad-CAM algorithm result when overlaying the

heat maps with the evaluated 100X magniﬁcation images.

with new images and information that will make it

possible to evolve the method presented in order to

classify images with cancer by the presence of HPV

and histological grade.

ACKNOWLEDGEMENTS

This work was supported by the Coordenac¸

de Aperfeic¸oamento de Pessoal de N

ıvel Superior

(CAPES) - Finance Code 001 and CAPES PDPG

Amaz

onia Legal 0810/2020 - 88881.510244/2020-

01; the Fundac¸

ao de Amparo

a Pesquisa e ao De-

senvolvimento Cient

ıﬁco e Tecnol

ogico do Maranh

(FAPEMA); and the Conselho Nacional de Desen-

volvimento Cient

ıﬁco e Tecnol

ogico (CNPq).

REFERENCES

ACS (2021). Tests for penile cancer - american cancer

society. https://www.cancer.org/cancer/penile-

cancer/detection-diagnosis-staging/how-

diagnosed.html. (Accessed on 10/06/2021).

Boumaraf, S., Liu, X., Zheng, Z., Ma, X., and Ferk-

ous, C. (2021). A new transfer learning based ap-

proach to magniﬁcation dependent and independent

classiﬁcation of breast cancer in histopathological im-

ages. Biomedical Signal Processing and Control,

63:102192.

Choudhary, T., Mishra, V., Goswami, A., and Sarangapani,

J. (2021). A transfer learning with structured ﬁlter

pruning approach for improved breast cancer classiﬁ-

cation on point-of-care devices. Computers in Biology

and Medicine, 134:104432.

Coelho, R., Pinho, J., Moreno, J., Garbis, D., Nascimento,

A., Larges, J., Calixto, J., Ramalho, L., Silva, A.,

Nogueira, L., Feitoza, L., and Barros-Silva, G. (2018).

Penile cancer in maranh

ao, northeast brazil: The high-

est incidence globally? BMC Urology, 18.

Cruz-Roa, A., Basavanhally, A., Gonz

alez, F., Gilmore, H.,

Feldman, M., Ganesan, S., Shih, N., Tomaszewski, J.,

and Madabhushi, A. (2014). Automatic detection of

invasive ductal carcinoma in whole slide images with

convolutional neural networks. In Gurcan, M. N. and

Madabhushi, A., editors, Medical Imaging 2014: Dig-

ital Pathology, volume 9041, pages 1 – 15. Interna-

tional Society for Optics and Photonics, SPIE.

Douglawi, A. and Masterson, T. (2017). Updates on the epi-

demiology and risk factors for penile cancer. Transla-

tional Andrology and Urology, 6:785–790.

Epstein, J., of Pathology, A. R., Magi-Galluzzi, C., Zhou,

M., Cubilla, A., of Pathology Staff, A. R., and

of Pathology (U.S.) Staff, A. F. I. (2020). Tumors of

the Prostate Gland, Seminal Vesicles, Penis, and Scro-

tum. AFIP atlas of tumor and non-tumor pathology.

Fifth series. American Registry of Pathology.

Filipczuk, P., Fevens, T., Krzy

zak, A., and Monczak, R.

(2013). Computer-aided breast cancer diagnosis based

on the analysis of cytological images of ﬁne needle

biopsies. IEEE Transactions on Medical Imaging,

32(12):2169–2178.

Huang, G., Liu, Z., Van Der Maaten, L., and Weinberger,

K. Q. (2017). Densely connected convolutional net-

works. In 2017 IEEE Conference on Computer Vision

and Pattern Recognition (CVPR), pages 2261–2269.

VISAPP 2022 - 17th International Conference on Computer Vision Theory and Applications

982

INCA (2021). Types of cancer — national can-

cer institute - jos

e alencar gomes da silva -

inca. https://www.inca.gov.br/tipos-de-cancer/cancer-

de-penis. (Accessed on 10/06/2021).

Kiani, A., Uyumazturk, B., Rajpurkar, P., Wang, A., Gao,

R., Jones, E., Yu, Y., Langlotz, C., Ball, R., Mon-

tine, T., Martin, B., Berry, G., Ozawa, M., Hazard, F.,

Brown, R., Chen, S., Wood, M., Allard, L., Ylagan,

L., and Shen, J. (2020). Impact of a deep learning

assistant on the histopathologic classiﬁcation of liver

cancer. npj Digital Medicine, 3.

Krizhevsky, A., Sutskever, I., and Hinton, G. E. (2012).

Imagenet classiﬁcation with deep convolutional neu-

ral networks. In Proceedings of the 25th Interna-

tional Conference on Neural Information Processing

Systems - Volume 1, NIPS’12, page 1097–1105, Red

Hook, NY, USA. Curran Associates Inc.

Kumar, A. and Shaik, F. (2015). Image Processing in Dia-

betic Related Causes. SpringerBriefs in Applied Sci-

ences and Technology. Springer Singapore.

Lecun, Y., Bottou, L., Bengio, Y., and Haffner, P. (1998).

Gradient-based learning applied to document recogni-

tion. Proceedings of the IEEE, 86(11):2278–2324.

Linkon, A. H. M., Labib, M. M., Hasan, T., Hossain,

M., and Marium-E-Jannat (2021). Deep learning

in prostate cancer diagnosis and gleason grading in

histopathology images: An extensive study. Informat-

ics in Medicine Unlocked, 24:100582.

Melo, R. C. N., Raas, M. W. D., Palazzi, C., Neves, V. H.,

Malta, K. K., and Silva, T. P. (2020). Whole slide

imaging and its applications to histopathological stud-

ies of liver disorders. Frontiers in Medicine, 6:310.

Morid, M. A., Borjali, A., and Del Fiol, G. (2021). A scop-

ing review of transfer learning research on medical

image analysis using imagenet. Computers in Biology

and Medicine, 128:104115.

Neto, A. D. (2012). Cytopathology technician: reference

book 3: Histopathology techniques — bras

ılia;

ministry of health of brazil ; 2012. 93 p.— ms.

https://pesquisa.bvsalud.org/bvsms/resource/pt/mis-

37140. (Accessed on 10/11/2021).

Rakhlin, A., Shvets, A., Iglovikov, V., and Kalinin, A. A.

(2018). Deep convolutional neural networks for breast

cancer histology image analysis. In Campilho, A.,

Karray, F., and ter Haar Romeny, B., editors, Im-

age Analysis and Recognition, pages 737–744, Cham.

Springer International Publishing.

Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S.,

Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bern-

stein, M., et al. (2015). Imagenet large scale visual

recognition challenge. International journal of com-

puter vision, 115(3):211–252.

Sarwinda, D., Paradisa, R. H., Bustamam, A., and Anggia,

P. (2021). Deep learning in image classiﬁcation using

residual network (resnet) variants for detection of col-

orectal cancer. Procedia Computer Science, 179:423–

431. 5th International Conference on Computer Sci-

ence and Computational Intelligence 2020.

Selvaraju, R. R., Cogswell, M., Das, A., Vedantam, R.,

Parikh, D., and Batra, D. (2019). Grad-cam: Visual

explanations from deep networks via gradient-based

localization. International Journal of Computer Vi-

sion, 128(2):336–359.

Sharma, H., Zerbe, N., Klempert, I., Hellwich, O., and Huf-

nagl, P. (2017). Deep convolutional neural networks

for automatic classiﬁcation of gastric carcinoma using

whole slide images in digital histopathology. Com-

puterized Medical Imaging and Graphics, 61:2–13.

Selected papers from the 13th European Congress on

Digital Pathology.

Spanhol, F. A., Oliveira, L. S., Cavalin, P. R., Petitjean, C.,

and Heutte, L. (2017). Deep features for breast cancer

histopathological image classiﬁcation. In 2017 IEEE

International Conference on Systems, Man, and Cy-

bernetics (SMC), pages 1868–1873.

Spanhol, F. A., Oliveira, L. S., Petitjean, C., and Heutte, L.

(2016a). Breast cancer histopathological image classi-

ﬁcation using convolutional neural networks. In 2016

International Joint Conference on Neural Networks

(IJCNN), pages 2560–2567.

Spanhol, F. A., Oliveira, L. S., Petitjean, C., and Heutte, L.

(2016b). A dataset for breast cancer histopathological

image classiﬁcation. IEEE Transactions on Biomedi-

cal Engineering, 63(7):1455–1462.

Srinidhi, C. L., Ciga, O., and Martel, A. L. (2021). Deep

neural network models for computational histopathol-

ogy: A survey. Medical Image Analysis, 67:101813.

Tajbakhsh, N., Shin, J. Y., Gurudu, S. R., Hurst, R. T.,

Kendall, C. B., Gotway, M. B., and Liang, J. (2016).

Convolutional neural networks for medical image

analysis: Full training or ﬁne tuning? IEEE Trans-

actions on Medical Imaging, 35(5):1299–1312.

Tellez, D., Litjens, G., B

andi, P., Bulten, W., Bokhorst, J.-

M., Ciompi, F., and van der Laak, J. (2019). Quanti-

fying the effects of data augmentation and stain color

normalization in convolutional neural networks for

computational pathology. Medical Image Analysis,

58:101544.

Thomas, A., Necchi, A., Muneer, A., Tobias-Machado, M.,

Tran, A. T. H., Van Rompuy, A.-S., Spiess, P. E., and

Albersen, M. (2021). Penile cancer. Nature Reviews

Disease Primers, 7(1):1–24.

Vieira, C., Feitoza, L., Pinho, J., Teixeira-J

unior, A., Lages,

J., Calixto, J., Coelho, R., Nogueira, L., Cunha, I.,

Soares, F., and Barros-Silva, G. (2020). Proﬁle of pa-

tients with penile cancer in the region with the highest

worldwide incidence. Scientiﬁc Reports, 10.

Vill

an, A. (2019). Mastering OpenCV 4 with Python: A

practical guide covering topics from image process-

ing, augmented reality to deep learning with OpenCV

4 and Python 3.7. Packt Publishing.

Wei, J., Tafe, L., Linnik, Y., Vaickus, L., Tomita, N., and

Hassanpour, S. (2019). Pathologist-level classiﬁca-

tion of histologic patterns on resected lung adenocar-

cinoma slides with deep neural networks. Scientiﬁc

Reports, 9.

Zhang, A., Lipton, Z. C., Li, M., and Smola, A. J.

(2021). Dive into deep learning. arXiv preprint

arXiv:2106.11342.

Zuiderveld, K. J. (1994). Contrast limited adaptive his-

togram equalization. In Graphics Gems.

Classiﬁcation of Histopathological Images of Penile Cancer using DenseNet and Transfer Learning

983