Prostate Cancer Detection, Segmentation, and Classiﬁcation using Deep

Neural Networks

Yahia Bouslimi, Takwa Ben A

ıcha Gader

and Afef Kacem Echi

National Superior School of Engineering, University of Tunis, LR: LATICE, Tunisia

Keywords:

Prostate Cancer, Computer-aided Diagnosis, Convolutional Neural Network, Magnetic Resonance Imaging,

MultiResU-Net, U-Net.

Abstract:

This paper provides a fully automated computer-aided medical diagnostic system that assists radiologists in

segmenting Prostate Cancer (PCa) Lesions from multi-parametric Magnetic Resonance Imaging (mp-MRIs)

and predicting whether those lesions are benign or malignant. For that, our proposed approach used deep

learning neural networks models such as residual networks (ResNet) and inception networks to classify clini-

cally relevant cancer. It also used U-Net and MultiResU-Net to automatically segment the prostate lesion from

mp-MRI’s. We used two publicly available benchmark datasets: the Radboudumc and ProstateX. We tested

our fully automatic system and obtained positive ﬁndings, with the AUROC of the PCa lesion classiﬁcation

model exceeding 98.4% accuracy. On the other hand, the MultiResU-Net model achieved an accuracy of

98.34% for PCa lesion segmentation.

1 INTRODUCTION

Prostate Cancer (PCa) is a common type of cancer

that affects the prostate gland, a small walnut-shaped

gland in the male reproductive system. It is the sec-

ond most common type of cancer in men worldwide,

with an estimated 1.4 million new cases diagnosed

each year, which is nearly (7.3%) of the total can-

cer cases worldwide.(Sung, 2021) The majority of

prostate cancer cases occur in men over the age of

65, and the risk of developing the disease increases

with age. The highest incidence rates of prostate can-

cer are found in North America, Europe, and East-

ern Asia, while it is less common in the other parts

of Asia and Africa.(global cancer observatory, ) De-

spite of the mortality rate for prostate cancer has been

decreasing in many countries due to improvements

in early detection and treatment. Current diagnostic

methods, such as digital rectal examination, PSA lev-

els, and biopsies, are imprecise (Nabil and Rahman,

2020). In most situations, those used in cancer screen-

ing procedures are insufﬁcient to effectively identify,

locate, and describe it, resulting in poor ultrasound

imaging quality. Furthermore, the biopsy samples

still just a group of random specimen that can be inac-

https://orcid.org/0000-0002-3786-3649

https://orcid.org/0000-0001-9219-5228

curate by missing their target. Multi-parametric Mag-

netic Resonance Imaging (mp-MRI) has emerged as

a promising area of medical diagnosis research in re-

cent years. It has shown promising results since its lo-

calization skills opened new avenues for critical PCa

detection, diagnosis, localization, risk stratiﬁcation,

and clinically signiﬁcant PCa staging. Despite the

increasing use of mp-MRI for PCa screening, there

remains a signiﬁcant need for uniformity in the in-

terpretation of mp-MRI images. Given the process’s

difﬁculty and intricacy, the radiologist frequently de-

pends solely on visual interpretation to complete the

work, which is complex. This could lead to inaccu-

racies and signiﬁcant inter-reader variability in diag-

nosis, especially when the mp-MRI sequences con-

tradict one other. When images contradict each other,

a Computer-Aided Decision-making System (CAD)

can reduce the rate of inaccurate diagnoses and pro-

vide better results. Given the use of deep learning ap-

proaches in biomedical image classiﬁcation and seg-

mentation, we propose a deep learning-based CAD to

assist radiologists in diagnoses and decision-making.

It is about a MultiResU-Net, an enhanced version of

the basic U-Net that allows it to run with fewer train-

ing sets while still providing accurate segmentation.

We will discuss the state-of-the-art PCa diagnostics-

based deep learning approaches in the upcoming sec-

tion. In the third section, we will describe and pre-

534

Bouslimi, Y., Gader, T. and Echi, A.

Prostate Cancer Detection, Segmentation, and Classiﬁcation using Deep Neural Networks.

DOI: 10.5220/0011795100003411

In Proceedings of the 12th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2023), pages 534-541

ISBN: 978-989-758-626-2; ISSN: 2184-4313

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

process the used data. We will present our proposed

deep learning model architectures in the fourth sec-

tion. Afterward, we will report our ﬁndings and com-

pare them to previous research. Finally, we conclude

by providing some insight into our future research.

2 RELATED WORKS

In the last few recent years, There has been many

research aiming to assist radiologists in appropri-

ately segmenting, locating, and identifying clinically

signiﬁcant prostate cancer. We can mention (Liu

et al., 2017) used XmasNet, a CNN-based classiﬁer

to categorize mp-MRI prostate cancer lesions on the

PROSTATEx dataset. With an AUC of 0.84, Xmas-

Net surpassed all typical machine learning models for

training and testing data. (Vente and Vos, 2021), on

the other hand, used 2D U-Net with MRI slices as

input and lesion segmentation maps that encode the

GGG By achieving a voxel-wise weighted kappa of

0.446% and a Dice similarity coefﬁcient of 0.37%,

the model beats traditional multi-class classiﬁcation

and multi-label ordinal regression. (Lehaire, 2016),

on the other hand, used their data, which included a

49-patient mp-MRI database and SVM-L and logis-

tic regression, to construct a CADe that generated a

cancer probability map for the radiologist. The draw-

back was that they built using the dictionary learning

method, and the dictionaries were approximated from

the features of previously retrieved images. Although

their results were statistically superior to the other

CADe diagrams, this difference needed to be more

evident on the sample probability maps. Meanwhile,

(Yunzhi Wang and Wang, 2018) demonstrated in

2018, the mp-MRI-based segmentation scheme out-

performs earlier T2W-based schemes by using a state-

of-the-art Fully Convolutional Network (FCN) archi-

tecture with residual connections to segment prostate

mp-MRI. (To et al., 2018), on the other hand, used a

3D deep, dense multi-path Convolutional Neural Net-

work (CNN) based on the encoder-decoder architec-

ture to segment the prostate in MRI images on two

distinct datasets. The encoder is composed of densely

connected layers, and the decoder interprets the fea-

tures and predicts the total prostate volume.

The developers of (Dai Z, 2020) used the

PROSTATEx dataset in their two-stage approach in

2019. They used a poorly supervised deep neural

network to detect and classify lesions after training

a Mask R-CNN model to segment the prostate struc-

tures automatically. On their validation set, this work

attained an average AUC of 0.912% and 0.882%.

(Zhenzhen Dai, 2019) employed Mask-RCNN for

prostate segmentation and Dominant Intra-prostatic

Lesion (DIL) segmentation and localization using

Mp-MRI in the same scenario. Furthermore, (Karimi

et al., 2019) trained two different CNNs to produce a

global CNN to create an automatic prostate segmenta-

tion method in T2-weighted images (T2w) in prostate

MRI sequence

Furthermore, in (Stahl, 2020), the authors used

a Genetic Algorithm to ﬁne-tune a trained CNN for

PCa detection to get a higher AUC on their 6-channel

diffusion-weighted prostate MRI dataset. On their

test dataset, this work yielded an AUC of 0.707. (Yoo

et al., 2019) also created and implemented an auto-

mated CNN-based process to detect clinically signiﬁ-

cant PCa in axial DWI images for each patient.

The authors of (Neto, 2020) published a deep

learning-based study of PCa via mp-MRI in 2020.

They experimented with numerous models, including

3D U-net, 3D Res-Net, and XmasNet. The outcome

was a DICE value of roughly 0.69, whereas other

researchers received a DICE score of 0.83. Mean-

while, (Liu, 2020) suggested a Pytorch V-Net deep

learning framework architecture on volumetric CNN

by borrowing the U-net network and dividing it into

residual stages of learning to conduct rapid and ac-

curate MRI prostate volume segmentation. (Nuhi

and Kevric, 2020) developed a novel PCa clinical

management prototype using nine classiﬁcation al-

gorithms in a PCa database. The classiﬁcation rate

for authors is 98.71%. For the AdaBoost classi-

ﬁer, sensitivity was around 97.4%, while speciﬁcity

was perfect and equivalent to 100%. Other algorithms

(Nave Bayes, Multi-layer Perception, Simple Logis-

tics, Nearest Neighbor, Random Committee, PART,

LMT, and Random Forest) also produced excellent re-

sults.

The authors of (Andrew Hwang, 2021) trained and

evaluated 13 different CNN architectures (binary clas-

siﬁcation models) on Raudboundumc MRI scans of

204 patients suspected of having prostate cancer, the

same one we utilized in our study. The best model

has an accuracy of 86.9% and an area under the ROC

(AUROC) of 90.3% percent. (B. Liu, 2021) pro-

posed an improved 2D U-Net model with an included

Squeeze-and-Excitation (SE) layer for prostate seg-

mentation using the public dataset PROMISE12 in

the same year (2021). The model is built around

an encoder stage that uses CONV blocks, SE lay-

ers, and max-pooling layers to extract features from

the input and a decoder stage that maps the returned

features to the original image. Experiments demon-

strated that the suggested model might outperform

other approaches in terms of segmentation accuracy

and DSC, with a mean DSC of 87%.

Prostate Cancer Detection, Segmentation, and Classiﬁcation using Deep Neural Networks

535

3 DATA PRE-PROCESSING

We used two separate datasets in this study. We ini-

tially used the Radboudumc dataset (see Table 1) for

PCa classiﬁcation. Following that, we used PRosta-

teX for PCa segmentation. The ProstateX dataset is

made up of 345 mp-MRI studies, and these studies

were split into training and test datasets. The train-

ing set consists of 204 mp-MRI studies, each involv-

ing one patient. Each study includes a T2w image

sequence, DW with b-values of b50, b400, and b800

s/mm2, an apparent diffusion coefﬁcient (ADC) map

(derived from the b-values), and Ktrans (computed

from dynamic contrast-enhanced-DCE-T1-weighted

series). The ProstateX2 challenge, which uses the

same data as ProstateX, should produce one to four le-

sion sites (points indicating their position on the mp-

MRI sequences) per patient; the average is 1.62. The

lesion’s location was documented under a competent

radiologist’s supervision. The test set consists of 140

mp-MRIs containing all previously speciﬁed informa-

tion.

Table 1: Detailed Radboudumc PCa Dataset Description.

Modalities MRI

Patients 204

Findings 326

Images before data augmentation 326

Images after data augmentation 1956

The only drawback with the Radboudumc dataset

was that it lacked lesion masks and needed cleaning

and cropping. Image cropping was required to re-

move unnecessary data (muscle, bones, tissues) and

leave only the prostate zone and its environs. Ini-

tially, we set the supplied lesion coordinate as our im-

age center; we took a 348 × 348 × 3 image as input;

we cropped it from all sides by 102 pixels, leaving

just an image of the prostate zone with a shape of

144 × 144 × 3. A similar procedure was performed

on ADC and Ktrans. The missing masks have to be

generated automatically. As a result, we extracted the

ROI (lesion) coordinates for each discovery ID num-

ber from the attached CSV data and grew a region of

5 × 5 white pixels in both height and width, where the

white box represents the mask. Then, to coincide with

the mp-MRI images, we constructed a 144 × 144 × 3

image with that box in the center. The result is shown

in ﬁgure 1, where (a) is an image of a cropped T2W

image and (b) is a cropped mask.

We also needed to undertake some data augmenta-

tion, but ﬁrst, we needed to integrate our mp-MRI se-

quences to visualize the lesion better. As a result, we

combined the T2W images with the ADC and Ktrans

(a) T2w MRI image (b) Result

Figure 1: mp-MRI used images.

MRI images. Figure 2 shows an example of the re-

sult, where (a) represents a cropped T2W image, (b)

and (c) are the matching ADC and Ktrans, and (d)

represents the merged image.

(a) T2w (b) ADC

Figure 2: mp-MRI used images.

We needed to preprocess the mp-MRI images for

the ProstateX dataset to clean the data and enrich the

images with extra information relevant to model train-

ing and testing. As a result, we cropped all T2w pho-

tos surrounding the prostate to have dimensions of

160 × 160 × 24 with a spacing of (0.5, 0.5, 3) mm.

Following that, we used third-order B-Spline inter-

polation for all image interpolation tasks, whereas

Gaussian label interpolation was used for segmenta-

tion masks. Only the location of the lesion center is

included in the ProstateX dataset. So we had to grow

them from the speciﬁed coordinates; the technique

is the same as in the Radboudumc dataset, except

that we utilized a threshold level set method from the

Python module SimpleITK72, which produces bet-

ter results and more visible boundary contours. We

thought about using prostate zone masks because the

radiologist could need such information.

ICPRAM 2023 - 12th International Conference on Pattern Recognition Applications and Methods

536

We also did some data augmentation on the

cropped and merged mp-MRI pictures to increase the

image quantity, improve our model performance, and

avoid over-ﬁtting. As a result, the data pool was ex-

panded by eleven-folds, and the number of images

and masks grew from 326 to 3586. After applying im-

age distortions, liquefying, and inward zooming tech-

niques, the zoom for each image and mask was ran-

dom (less or equal to 50%). We also used horizon-

tal and vertical mirroring and random rotations up to

20 degrees off their original axes. Figure 3 depicts

an example of the data augmentation procedure on an

image. 3(a) and (b) are original images of a merged

T2W, ADC, and Ktrans sequence, whereas (c) and

(d), are examples of augmented images. We randomly

divided the dataset into training and testing datasets

in 1 to 7 ratios once it was complete. Consequently,

we obtained 3075 photos (or 85.72%) for training and

513 images (or 14.28%) for testing.

(a) Original (b) Original copy

Figure 3: Example of augmented images.

We took a different strategy with ProstateX. To

correctly apply some signiﬁcant 3D data augmenta-

tion, we applied rigid and non-rigid transformations

derived from scaling, rotations, and elastic deforma-

tions. In terms of data splitting, the ProstateX design-

ers have already divided it into train and test datasets.

4 PROPOSED DEEP MODELS

We used two separate model groups. The ﬁrst in-

cludes the ResNet50, ResNet101 V2, and ResNet152

V2 models. We used it to differentiate between se-

vere and benign lesions in the PCa classiﬁcation chal-

lenge. For the PCa segmentation or semantic seg-

mentation of PCa lesions, we used the second model

group, which includes the U-Net and MultiResUnet

models.

4.1 PCa Classiﬁcation

For PCa classiﬁcation, we attempt to determine

whether or not there is a malignant lesion in the sup-

plied prostate mp-MRI image. To construct a good

classiﬁer, we used deep learning models. We tested

four models to see which would produce the most sig-

niﬁcant results: we used InceptionV3, the ResNet50,

the ResNet101 V2, and the ResNet152 V2. Compared

to other models, Both InceptionV3 and ResNet50

(see ﬁgure 4)produced the most promising results.

The ResNet50 architecture enables large numbers of

convolutional layers to function efﬁciently. Adding

many deep layers in ResNet50, ResNet101 V2, and

ResNet152 V2 has signiﬁcant drawbacks; for exam-

ple, they generate vanishing gradient difﬁculties, re-

sulting in a degradation of output and training perfor-

mance.

4.2 PCa Segmentation

For PCa segmentation, we used an mp-MRI image

(DICOM format) as input and performed automatic

seamless prostate lesion segmentation. We saved

computer resources by decompressing and transform-

ing DICOM data into a more legible format: a series

of two-dimensional images in PNG format. Then we

cut their size by half (making all photos (128 × 128)

in size). Finally, the proposed system uses the trained

segmentation model (MultiResU-Net) to segment and

locate the prostate lesion in the input images.

The MultiResU-Net architecture is inspired by the

U-Net architecture, as seen in ﬁgure 5. As a re-

sult, it has the symmetrical shape of a ’U’, and it is

made up of an encoder, a decoder, and a MultiRes-

Path that relies on them. The encoder extracts the

main features from the input images before passing

them to the decoder, which creates the corresponding

segmentation map from the extracted features. As for

MultiResPaths, they have upgraded skip connections

that rely on the encoder and decoder together. All

these components have been collected and combined

in MultiResU-Net to form MultiResBloc, which sub-

stitutes the skip connection.

The MultiRes block not only substitutes the U-

three Net’s convolutions but also outperforms them

by learning additional spatial information. It was con-

structed on the ResNet architecture. This skip link is

helpful in biomedical image segmentation. Jump be-

tween convolutional layers to mitigate, if not elimi-

Prostate Cancer Detection, Segmentation, and Classiﬁcation using Deep Neural Networks

537

Figure 4: Used ResNet50 architecture.

nate, the vanishing gradient problem induced by nu-

merous deep CNN layers, as seen in ﬁgure 5. Fur-

thermore, the MultiRes block gradually increases the

number of ﬁlters to reduce memory loss, whereas the

standard U-Net architecture uses an equal number of

ﬁlters on each layer. MultiResU-Net gradually raised

the number of ﬁlters from 3 × 3, 5 × 5 to 7 × 7 and

concatenated them using the residual connection of

1 × 1 ﬁlter to conserve dimensions while learning and

avoiding the vanishing gradient problem.

MultiResU-Net transformed U-Net by introduc-

ing shortcut connections between convolutional lay-

Figure 5: MultiResUNet block architecture.

ers before and after max-pooling (skip connections).

During the pooling procedure, the latter allows the

network to send spatial information from the encoder

to the decoder. The skip connections are subjected

to a number of non-linear modiﬁcations. To facilitate

learning, these transformations are convolutional lay-

ers paired with residual connections.

Table 2: MultiResU-Net Models Hyperparameter.

Hyperparameter MultiResU-Net

Train images 1564

Validation images 392

Epochs 50

Batch size 30

batch size 15

Alpha 1.36

Optimizer ‘adamax’

Learning rate 0.0005

Beta 1 0.8

Beta 2 0.888

Epsilon 1e-07

Loss binary crossentropy

As a result, the feature is subjected to a convo-

lutional layer sequence. These two are linked via

residual links. The output will then be concatenated

with the decoder features. Our model architecture

is depicted in 6. Its overall architectural shape is

highly similar to that of U-Net. It comprises 13

Multires-Blocks layers, replacing the traditional U-

Net sequences with two convolutional layers. To

make the model more resilient and reduce over-ﬁtting,

we added a 30% dropout function to each multires-

block. Additionally, we eliminated the skip connec-

tions of the U-casual Net in favor of ResPaths. The

same is true for ResPaths, which have a 30% dropout

function. Table 2 displays a detailed description of

the parameters used during the training process.

5 EXPERIMENTATION AND

RESULTS DISCUSSION

5.1 PCa Classiﬁcation

We aim to assess whether each mp-MRI shows a

clinically signiﬁcant cancer or a benign lesion. It is

a straightforward binary classiﬁcation; initially, we

ICPRAM 2023 - 12th International Conference on Pattern Recognition Applications and Methods

538

Table 3: Used classiﬁcation models results.

Model Accuracy Precision Speciﬁcity Recall/Sensitivity AUROC

InceptionV3 96.3% 90.8% 98% 92.5% 98.4%

ResNet50 94.7% 83.4% 95% 93.5% 98%

ResNet101V2 92.6% 80% 92.65% 92.3% 92.2%

ResNet152V2 90.3% 68% 91.46% 84.61% 90.5%

Figure 6: MultiResU-Net model architecture (Nabil and

Rahman, 2020).

picked U-Net because it is one of the ﬁnest models for

processing biomedical images ranging from mp-MRI

to ultrasound and microscopic images. However, be-

cause it was designed for semantic segmentation, it

still has numerous limitations in mp-MRI image cat-

egorization. As a result, we attempted to predict PCa

from mp-MRI using a distinct variety of pre-trained

ResNet architectures and Inception-V3 CNNs. Our

models were trained across 150 epochs.

The ResNet50 performed best at epoch 62, while

InceptionV3 performed admirably at epoch 149.

ResNet50 beats both ResNet101-V2 and ResNet152-

V2 in all ﬁelds, as shown in Table 3. It has a higher

area under the receiver operating characteristic (AU-

ROC) score (98%) than its siblings ResNet101V2

and ResNet152-V2, which had 92.2% and 90.5%, re-

spectively. The same is true for accuracy, precision,

and sensitivity, which ResNET50 has scored sequen-

tially (94.7%, 83.44%, 95.4%, and 93.5%), outper-

forming ResNet101V2 scores. It is also worth noting

that this latter outperformed ResNet152-V2 in preci-

sion and recall (80% vs. 68% for precision and 92.3

vs. 84.61%) and slightly beat it in the other metrics.

Inception-V3, on the other hand, surpassed ResNet50

in all metrics except recall, which was lower than ex-

pected from Inception-outstanding V3’s performance.

Our top model got 96.3%, 98%, and 98.4% for accu-

racy, precision, and AUROC, which is much higher

than most achieved outcomes in the ﬁeld.

The table 4 show the AUROC curves results. The

four models produce good results, above 90%, and

the InceptionV3 gives nealy perfect result of 98.4%.

They also gave a noisy loss and accuracy curves in the

beginning that leans to be stedy over time. ResNet50

outperforms his siblings in the learning rate, gradually

increasing with wavy movements to achieve high ac-

curacy in the training and testing phases. in our case,

InceptionV3 surpassed ResNet50 by a small perse-

nage and shows a good ﬁt of training and validation

loss curves. Similarly, the other models provide a

good ﬁt; however, the curves were slightly noisier

than ResNet50’s.

5.2 PCa Segmentation

Reacently, many attempts have been undertaken to

auto-segment Prostate cancer utilizing deep learning-

based approaches. The research ranges from seg-

menting the prostate’s normal zonal structure to PCa

lesions, but an agreement has yet to be reached on the

combination of input sequences of mp-MRI images.

In this study, we used MultiResU-Net to automati-

cally generated Lesion masks from mp-MRI’s. We

found that the three recognized models behaved simi-

larly. The indicated models in Table 5 showed equiv-

alent results when segmenting the PCa lesions from

the PROSTATEx dataset (DSC of 52.73, 41, 37.46

percent consecutively for (Lai and Shen., 2021), (?),

(Vente and Vos, 2021)). The Modiﬁed-SegNet model

produced consistently improved results. Our model

performance, however, outperformed this with a Dsc

of 59%. Previous studies that used simple T2W im-

ages for training came to the same conclusion. To

reach this result, we combined our T2W, ADC, and

Ktrans MRI images as three channels of an RGB im-

age. Regarding test accuracy ﬁndings, shown in Ta-

ble 5, our employed model beats the modiﬁed SegNet

model presented in (Lai and Shen., 2021). The latter

has an accuracy of 96.97%, whereas our model has

an accuracy of 98.14% and a loss of 0.074%. Simi-

larly, our model beat other models in terms of AUC.

We received a score of 74%, compared to the 90%,

94%, and 84% obtained by (Lai and Shen., 2021)

((Song, 2018), (Vente and Vos, 2021)). Furthermore,

the ground truths are generated automatically rather

than delivered by an expert. As a result, the train-

ing masks were more general than the generated ones;

they differed in shape and form and were particular to

the lesion, unlike the automatically created mask.

Prostate Cancer Detection, Segmentation, and Classiﬁcation using Deep Neural Networks

539

Table 4: Prostate cancer classiﬁcation curve results.

AUROC CURVE ACCURACY/LOSS CURVE

InceptionV3 AUROC curve InceptionV3 accuracy/loss curve

AUROC CURVE ACCURACY LOSS CURVE

Table 5: MultResU-Net Performance Comparison.

Author Model Accuracy Loss DSK AUROC

Our work MultiResU-Net 98.34% 0.074% 59% 74%

(P.Valero and et al., 2021) Retina-U-Net 100% - 79% 95%

(P.Valero and et al., 2021) Retina-U-Net 100% - 37.5% 87%

(Lai and Shen., 2021) Modiﬁed-SegNet 96.97% - 52.73% 90%

(Vente and Vos, 2021) U-Net - - 37.46% -

(P.Valero and et al., 2021) Retina-U-Net - - 100% 87%

(Song, 2018) VGG-Net - - - 94%

(Liu et al., 2017) XmasNet - - - 84%

ICPRAM 2023 - 12th International Conference on Pattern Recognition Applications and Methods

540

6 CONCLUSION

A CAD for fully automatic Pca segmentation, and

classiﬁcation of mp-MRI images was reported in

this study. We tested it using two publicly avail-

able benchmark databases: the RadboudUMC dataset

and the ProstateX dataset. Our proposed system has

demonstrated high efﬁciency in all the key operations

for which it was designed. It achieved a classiﬁcation

accuracy of 96.3% and an AUROC of 90% with In-

ceptionv3 model. Likewise, for segmentation using

MultiResUnet, it achieved 98.34% percent accuracy

with a 74% AUC, which can still be improved with a

few more tweaks. The proposed models could learn

more global and local features and produce better re-

sults. We are now working on determining the PCa

Gleason grade using mp-MRI. In future work, we aim

to improve the performance of the proposed model.

We also intend to test the model on other benchmark

databases and use it to detect and segment different

cancer kinds to demonstrate the approach’s generaliz-

ability.

REFERENCES

Andrew Hwang, Emlyn Evans, E. M. J. L. S. J. . T. L.

(2021). Research report on deep learning methods in

prostate cancer.

B. Liu, J. Zheng, P. C. S. L. C. L. (2021). An improved 2d u-

net model integrated squeeze-and-excitation layer for

prostate cancer segmentation.

Dai Z, Carver E, L. C. e. a. (2020). Segmentation of the Pro-

static Gland and the Intraprostatic Lesions on Multi-

parametic Magnetic Resonance Imaging Using Mask

Region-Based Convolutional Neural Networks., pages

5(3):473–481.

global cancer observatory. Cancer today.

Karimi, D., Samei, G., Shao, Y., and Salcudean, S. E.

(2019). A deep learning-based method for prostate

segmentation in t2-weighted magnetic resonance

imaging. ArXiv, abs/1901.09462.

Lai, Chih-Ching, H.-K. W. F.-N. W. Y.-C. P. T.-P. L. H.-

H. P. and Shen., S.-H. (2021). Autosegmentation of

prostate zones and cancer regions from biparametric

magnetic resonance images by using deep-learning-

based neural networks.

Lehaire, J. (2016). D

etection et caract

erisation du cancer de

la prostate par images irm 1.5t multiparam

etriques.

Liu, S., Zheng, H., Feng, Y., and Li, W. (2017). Prostate

cancer diagnosis using deep learning with 3d multi-

parametric mri.

Liu, Y. (2020). 3d image segmentation of mri prostate based

on a pytorch implementation of v-net.

Nabil, I. and Rahman, M. S. (2020). Multiresunet: Rethink-

ing the u-net architecture for multimodal biomedical

image segmentation.

Neto, P. (2020). Deep learning based analysis of prostate

cancer from mp-mri.

Nuhi

c, J. and Kevric, J. (2020). Prostate Cancer Detection

Using Different Classiﬁcation Techniques, pages 67–

73.

P.Valero, O. J. and et al., J. L. (2021). Deep learning for

fully automatic detection, segmentation, and gleason

grade estimation of prostate cancer in multiparametric

magnetic resonance images.

Song, Y.; Zhang, Y. (2018). Computer-aided diagnosis of

prostate cancer using a deep convolutional neural net-

work from multiparametric mri.

Stahl, B. (2020). deepSIP: deep learning of Supernova

Ia Parameters. Astrophysics Source Code Library,

record ascl:2006.023.

Sung, H. e. a. (2021). “global cancer statistics 2020: Globo-

can estimates of incidence and mortality worldwide

for 36 cancers in 185 countries.”.

To, M., Vu, Q., Turkbey, B., Choyke, P., and Kwak, J. T.

(2018). Deep dense multi-path neural network for

prostate segmentation in magnetic resonance imaging.

International Journal of Computer Assisted Radiology

and Surgery, 13.

Vente, C. and Vos, P. (2021). Deep learning regression for

prostate cancer detection and grading in bi-parametric

mri.

Yoo, S., Gujrathi, I., Haider, M., and Khalvati, F. (2019).

Prostate cancer detection using deep convolutional

neural networks.

Yunzhi Wang, Bin Zheng, D. G. and Wang, J. (2018).

Fully convolutional neural networks for prostate can-

cer detection using multi-parametric magnetic reso-

nance images: an initial investigation.

Zhenzhen Dai, Eric Carver, C. L. J. L. A. F.-W. Z. M. P. M.

E.-N. W. (2019). Segmentation of the prostatic gland

and the intraprostatic lesions on multiparametic mri

using mask-rcnn.

Prostate Cancer Detection, Segmentation, and Classiﬁcation using Deep Neural Networks

541