Classification of prostate cancer using Deep Learning approach and MobileNetV2 architecture

doi:10.21203/rs.3.rs-1964155/v1

Since prostate cancer is one of the most important causes of mortality in today's society, the study of why and how to diagnose and predict them has received much attention from researchers. The collaboration of computer and medical experts offers a new solution in analyzing this data and obtaining useful and practical models, which is data mining. In fact, data mining, as one of the most important tools for data analysis and discovering the relationships between them and predicting the occurrence of events is one of the practical tools of researchers in this way. This study diagnoses and classifies prostate cancer using Deep Learning approach and MobileNetV2 architecture based on a method to identify the factors affecting this disease. In this study, data was taken from a database on the Brigham Hospital website. Also, in order to improve the methods of diagnosing prostate cancer, a feature-classification approach has been proposed, which has been evaluated using a data set related to clients' files. The proposed method after applying various classification methods on the available data including benign and malignant diagnosis and reaching an optimal method with relatively high accuracy using a faster R-CNN network to segment the area and later using architecture Various convolutional neural networks (CNNs) have been selected for feature extraction and set classification, increased processing speed. In addition, the MobileNetV2 architecture is used, which has the ability to achieve AUC in the range of 0.87 to 0.95 with acceptable performance, high processing speed and relative accuracy for the diagnosis of prostate cancer.

MobileNetV2 architecture

various convolutional neural networks (CNN

prostate cancer prediction

processing speed

Prostate cancer is common cancer among men and is one of the leading causes of death. Statistics for 2018 indicate an equal number of 1.276106 new cases of prostate cancer, suggesting the untimely and inaccurate diagnosis of this disease and its symptom[1]. Prostate cancer is asymptomatic during its early stages and can be expressed only via PSA (prostate-specific antigen)[2].

With PSA increasing, tissue sampling is also carried out that may prove the presence of cancer. But to determine the progression of the disease and its stage, it is necessary to use imaging. The reason is that it is much more reliable and less repeatable; it is also less dangerous than sampling (biopsy)[3].

Magnetic imaging by providing an acceptable report on the condition of prostate and PI-RADS[4] increase sensitivity in diagnosing prostate cancer. As there are various magnetic imaging techniques such as T2W, T1W, DWI, and ADC, a correct diagnosis is also highly a matter of the radiologist's experience and physician[5].

Figure 1 displays an image of prostate MRI scanned using T2W. Both images show prostate cancer, but as it is seen, the malignant tumor type is dark, and the benign one is light. In the images provided by T2W, prostate cancer tends to be seen as an area of low signal in an area of high signal[6].

Hence, the accuracy in imaging and choosing features effective in the disease are very important. Deep learning techniques which use extremely powerful convolutional neural networks have been hugely successful in diagnosing and classifying lesions in medical images[7].

But what has called these techniques into question is the limited data they provide for these networks as the input for training. This article has aimed to present a new method for increasing data. Then, results have been examined with the application of a deep learning network for diagnosing prostate cancer. Due to the high importance of morphological features of images, this method enables the production of new images with high differentiation. In the end, performance has been assessed by comparing the results obtained from other methods and those from this method. In [8]deep learning network and protocol T2W is used, in this article the u-net architecture is combined with the ResNet50 architecture and finally to separate benign prostate cancer from malignant, the result is AUC = 0.645 Is. In [9] with a change in the ResNet architecture and the use of images with DWI protocol related to MRI and forest classification to finally separate benign prostate cancer from malignant, the result is AUC = 0.84. in [10] With several different protocols, prostate cancer examination by a physician such as DRE and PSA and MRI images and the use of deep learning networks resulted in AUC = 0.889. in [11] reviewed articles on the use of CNNs in the diagnosis of prostate cancer and concluded that the mean AUC was 0.79 (range between 0.77 and 0.87).

The images used in this study are taken from a database on the Brigham Hospital website[12]. In this database, MRI images of 230 patients of different weights are available. Of course, a small number of these images were usable in image processing, so the number of images approved by the doctor with the method has been increased. All patients registered on this site have prostate cancer and the benign and malignant type of this disease has been confirmed by a reputable doctor. In this study, the number of MRI images of the prostate using T2W falling into two benign and malignant types by the physician increased using the popular methods of noise removal in image processing. Then to separate the prostate gland from the rest of the image, the efficient method of Faster R-CNN utilized for region segmentation. Later, for feature extraction and automatic classification, a convolutional neural network with different architectures was used. Finally, these architectures will compare in terms of their efficiency. The results will compare with that of similar methods, which showed the efficiency of this method in response speed and relative accuracy. Figure 2 displays a general overview of the presented methods.

2.1 Data Preprocessing and Augmentation

In this study, MRI images were used using T2W related to 31 prostate cancer patients. This number of images was too small for an in-depth learning-based education system. Thus, it is possible to increase this number using common techniques of data augmentation like rotation, scaling, cropping, translation, color augmentation and then carry out preprocessing which includes applying noise removal filters[1, 13]. Here to minimize the execution time, data augmentation was done using the popular methods of image processing. Due to its valuable characteristics in extracting the features of images, morphology was also used for increasing their number. Structural and formal characteristics like shape, color, and patterns are investigated in morphology. The basic morphological operators are erosion, opening, closing, and dilation. These operators are used for filling holes, connecting objects, etc. In morphological processing, the sizes of the input and output images are the same. In utilizing the dilation operator, the pixels related to the borders of objects are added. In erosion, the pixels of the borders of objects are removed. The closing and opening operators are a combination of the other two operators. Due to their characteristics, the dilation and erosion operators are more widely used[14].

And this article has used these two more common operators, too.

Gaussian filters[15], histogram equalization[16], and the median filter are among popular filtering methods in image processing for removing noises common in MRI images[17, 18].

Hence, we will have more realistic images for processing. Surely the response speed will go higher with the removal of the preprocessing stage independently and the combination of the preprocessing stage and data augmentation. Therefore, in addition to the image itself, five different types of masks have been applied to the images, bringing their number to 186.

2.2 Separating

Zoning or separating the image means separating one part of the image from the others. Separating one area from the rest of the image can be done using area segmentation methods. Currently, one of the popular methods in this regard is Faster R-CNN[19]. This network has several advantages for region segmentation; in this network, the features of each convolutional layer can be used for predicting the related region proposals. Hence, there is no need for an algorithm like selective search, and therefore the model gains speed without repetitive computations. Faster R-CNN is formed of two modules: RPN [20]and Fast R-CNN[21]. In this network, the input image enters the RPN network after mapping the feature by CNN network convolution layers. Suggested areas are generated in RPN; these areas are likely to be present in the area in question. This network has many advantages over one-stage systems such as Yolo [22], SSD[23]. Among other things, the computational cost of the proposed areas is significantly reduced and also this network has no limit to perform calculations on the size of the input image and it can be implemented on real images.

Figure 3 shows how the Faster R-CNN network works on the image. The extraction process of the prostate area is depicted. After extracting the specific features through the domain selection network, the domains produce the same size called ROI, then transfer to the fully connected layer, where they are sorted by a Fast R-CNN layer.

In this research, 80% of images are given to Faster R-CNN as the training. This training is done by the pre-trained network AlexNet[24] then 20% of images are used for assessing the network[25, 26]. In this way, the prostate area is separated from the rest of the MRI as the target area.

2.3 Feature Extraction and Classification by Deep Learning Network

In recent years, deep learning networks have helped physicians diagnose diseases. Using improved algorithms, the method of deep learning can provide precise and professional predictions about an image. This is carried out by creating a training model from a separate data set, the validity of which is later assessed. In unsupervised learning, the software identifies a behavioral algorithm unassisted. In supervised learning, a new set with correct tags is produced. Today machine learning systems are mostly based on supervised learning. In these systems, pre-trained learning networks are used. In this method, each layer is first trained independently and then is done on a precisely regulated and integrated network[27]. Some features need shallow networks while others produce better responses with deeper networks. However, the main aim of medical image processing is to segment images into several regions having the same features. In this regard, CNNs are best for this challenge because feature extraction and classification are done simultaneously in them. In this research, different CNN architectures were compared to diagnose the two types of benign and malignant prostate cancers[18].

Much work has been done to use CNN[28] to diagnose prostate cancer and the extent of the disease[19]. In this study, different CNN architectures for the diagnosis of prostate cancer in both benign and malignant types are compared, and the results show the acceptable performance of the proposed design.

2.3.1 ResNet Neural Network

Residual neural networks rank first in regard to the challenge of LSVRC2015. ResNet architecture has been designed with a maximum of 152 layers for training and learning. This architecture outperforms its previous architectures like AlexNet and GoogleNet[29]. In this structure, the network is formed of accumulated residual blocks. Different architectures with various layers have been introduced so far including ResNet18 where 18 layers are used for training and learning. ResNet34, ResNet50 and ResNet101 are 34, 50 and 101 layers deep, respectively[30]. The more layers, the deeper the network, and the better the response should be. But if the response speed is lower and the relative accuracy is desired, shallow networks should also be considered.

2.3.2 MobileNetv2 Neural Network

It is a convolutional neural network that has 53 layers of depth. You can download a pre-training version of the trained network of over one million images from the ImageNet database. The pre-trained grid can categorize images into 1000 categories of objects, such as keyboards, mice, pencils, and many animals. The MobileNetV2 architecture reverses from a residual structure in which the input and output of the remaining blocks are thin layers of bottlenecks. It also uses lightweight twists to filter the properties of the expansion layer. Finally, it eliminates nonlinearities in thin layers. High speed, low number of parameters, and acceptable accuracy feature strongly in MobileNetv2 neural networks. Having a convolution different from the standard one is the reason for having the minimum parameter in this architecture[31].

The internal structure of the CNN network is shown in Fig. 4. This figure shows how to extract features and classify data. In this way, the images from the previous stage, after passing through the convolution layers and pooling layers, the features extracted by the fully connected layers are divided into two categories, benign and malignant.

2.4 Performance Assessment Criteria

One of the main stages after designing the proposed model is performance assessment, accuracy, and truth. Sensitivity and specificity are two vital indices for the statistical evaluation of performance and the results of classification tests. The quality of the proposed method is possible to be measured and described using sensitivity and specificity[32]. Post-analysis data are classified as below:

True positive (TP), false positive (FP), true negative (TN), and false-negative (FN).

Figure 5 shows the location of the TP, FP, TN, and FN criteria in the confusion matrix.

In this research, malignancy has been chosen as the first group and tagged as positive. Benignancy has been selected as the second group and labeled as negative. Thus if TP is close to 100%, malignant tumors have been correctly diagnosed and if FP is big, the group wrongly diagnosed as malignant has a larger percentage. In order to simultaneously show all criteria, a confusion matrix has been drawn. This matrix is usually used for supervised algorithms. The criteria derived from the confusion matrix are calculated as follows:

Accuracy indicates the correct diagnosis for the whole set, which is calculated by formula 1.

Accuracy= (TP + TN)/(TP + TN + FP + FN) (1)

Sensitivity shows the correct diagnosis of the first group, which is calculated by formula 2.

Sensitivity = TP/(TP + FN) (2)

Specificity shows the correct diagnoses of the second group, which is calculated by formula 3.

Specificity = TN/(TN + FP) (3)

Another criterion used for the measurement is the receiver operating characteristic curve (ROC) where the vertical axis is the true rate of change for the first group (true positive) and the horizontal axis is the false rate of change in the first group (false positive)[33].

One more criterion utilized for performance assessment in machine learning methods is the area under the curve (AUC). AUC value obtains ROC, showing a number between 0 and 1. The closer this number is to 1, the truer diagnoses are[34].

In this section, the results obtained from the proposed method have been given. The number of images has been increased to 186 and in the first stage, the region of interest (prostate gland) has been separated from the rest of the MRI. Then, image classification into two benign and malignant types has been done using CNN with different architectures. The total data was divided into 80 to 20, meaning that 80% of the images were for training and 20% for experiments, validation data is used to adjust the model parameters. And since the validation data should not be the same as the training data therefor the validation data should not be used in the training phase, so 20% of the training images are used for validation. These architectures have been ResNet18, ResNet50, ResNet101, and MobileNetV2, the performances of which have been compared in regard to two malignant and benign types.

It is clear that the best classification performance in the ROC diagram is achieved when the classifier shows a point (0, 1), and this is exactly the case when the classifier has the lowest error rate and the highest evaluation or sensitivity rate.

In Fig. 5 shows the ROC diagram for the proposed method by ResNet18 and mobilenetv2.

In Fig. 6 shows the ROC diagram for the proposed method by ResNet50 and ResNet101.

The diagrams generated in ROC accurately show that the performance stability is higher in the ResNet50 architecture, which corresponds to the results obtained for the AUC criterion.

In Table 1, the Accuracy criterion for four architectures) ResNet18, ResNet50, ResNet101 and MobileNetV2 (has been calculated.

Table 1

Results for Accuracy criteria with four selected methods: ResNet18, ResNet50, ResNet101 and MobileNetV2.
Type of architecture in deep learning network	Accuracy
ResNet101	86%
ResNet50	86%
ResNet18	86%
MobileNet_V2	82%

Table 2, AUC criteria compare the four selected architectures.

Table 2

Results for AUC criteria with four selected methods: ResNet18, ResNet50, ResNet101 and MobileNetV2.
Type of architecture in deep learning network	AUC
ResNet101	0.91
ResNet50	0.95
ResNet18	0.94
MobileNet_V2	0.87

In Table 3, Specificity and sensitivity criteria were calculated for the four selected architectures( ResNet18, ResNet50, ResNet101 and MobileNetV2. )

Table 3

Results for specificity and sensitivity criteria with two selected methods) ResNet18, ResNet50, ResNet101 and MobileNetV2.
Type of architecture in deep learning network	Specificity(%)	Sensitivity(%)
ResNet101	78.94	90.32
ResNet50	73.68	93.54
ResNet18	89.47	83.87
MobileNet_V2	94.73	74.2

The computer execution of the recommended method has been conducted in MATLAB with GPU.

The execution time of this method using ResNet18, ResNet50, ResNet101, and MobileNetV2 architectures has been shown in Table 4.

Table 4

the duration of the proposed program with the selected architectures.
Type of architecture in deep learning network	Running time
ResNet101	84 min,43 sec
ResNet50	53 min, 5 sec
ResNet18	18 min,59 sec
MobileNet_V2	29min, 24 sec

As it is seen, the execution time in ResNet18 has given the best response. MobileNetv2 has also, however, an acceptable response.

The results clearly indicate that response accuracy is the same in three ResNet architectures and a little lower in MobileNetv2 architecture. Hence, if a general accuracy is desired, each of these architectures can be a correct choice. But given the lower response time in ResNet18, this architecture is the better choice. And if specificity is regarded as the selection criterion, MobileNetv2 architecture will provide a much better response than other ones. And should sensitivity be seen as the criterion, ResNet50 will be an appropriate option. And in the end, due to having the lowest number of parameters for training and relative accuracy, MobileNetv2 can be the most appropriate choice in all fields.

the comparison of the AUC results obtained from the proposed method with the mobilenetV2 architecture with the methods that worked in this field in recent years is shown in Table 5.

Table 5

Comparison of the obtained result AUC with the proposed method with the result obtained in recent years.
Methods	Year of publication	AUC
J. Ishioka et al,[24].	2018	0.645
S. Yoo,[25].	2019	0.84
P. Woźnicki et al[26].	2020	0.889
J. M. T. Castillo, [27].	2020	0.79
The proposed method with mobilentv2		0.87

These comparisons show that the accuracy obtained by the proposed method for medical applications and initial studies can be acceptable.

The present research aimed to segment and classifies prostate cancer into two types, benign and malignant, based on MRI images. And to be able to arrive at an appropriate answer by classifying images based on deep learning networks. And of course, improving the training process by increasing the number of images using noise removal filters according to the noise in MRI images and use very effective morphological features in images, such as very important data in image processing. The results show that relative accuracy and acceptable performance speed of the low-parameter MobileNetV2 architecture compared to the three ResNet-18, ResNet-50, and ResNet101 architectures can reflect the appropriate performance of this processing scheme.

Availability of data and materials

The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.

Declaration of Competing Interest:

The authors declare that they have no known competing financialinterests or personalrelationships that could have appearedto influence the work reported in this paper.

Funding:

Funding Name: Shahid Chamran University of Ahvaz, Funding Number: 1.

Acknowledgments: The work described in this paper was supported by the Shahid Chamran University of Ahvaz, as a PhD. The authors would like to thank the Shahid Chamran University of Ahvaz for financial support.

CRediT authorship contribution statement:

The N.Pirzad-Mashak conceived the study, participated in its design and coordination, and helped produce the manuscript. The G. Akbarizadeh and E. Farshidi participated in the design and coordination and helped draft the version. All authors of the final version have read and approved the manuscript and agree with it.

Consent for publication:

All the authors have consented to the publication of this manuscript.

Competitive Advantages:

The authors participating in the study stated that they do so ؛do not have any conflict of interest with this version.

Author details:

¹Department of Electrical Engineering, Mahshahr Branch, Islamic Azad University, Mahshahr, Iran.

²Department of Electrical Engineering, Faculty of Engineering, Shahid Chamran University of Ahvaz, Ahvaz, Iran.

Shorten, C. and T.M. Khoshgoftaar, A survey on image data augmentation for deep learning. Journal of big data, 2019. 6(1): p. 1–48.
Mayerhoefer, M.E., et al., Effects of MRI acquisition parameter variations and protocol heterogeneity on the results of texture analysis and pattern discrimination: an application-oriented study. Medical physics, 2009. 36(4): p. 1236–1243.
Klotz, L., Contemporary approach to active surveillance for favorable risk prostate cancer. Asian journal of urology, 2019. 6(2): p. 146–152.
Purysko, A.S., et al., RadioGraphics Update: PI-RADS Version 2.1—A Pictorial Update. Radiographics, 2020. 40(7): p. E33-E37.
Cheung, D.C. and A. Finelli, Magnetic resonance imaging diagnosis of prostate cancer: promise and caution. Cmaj, 2019. 191(43): p. E1177-E1178.
Hambarde, P., et al., Prostate lesion segmentation in MR images using radiomics based deeply supervised U-Net. Biocybernetics and Biomedical Engineering, 2020. 40(4): p. 1421–1435.
Gaunay, G., et al., Role of multi-parametric MRI of the prostate for screening and staging: Experience with over 1500 cases. Asian journal of urology, 2017. 4(1): p. 68–74.
Ishioka, J., et al., Computer-aided diagnosis of prostate cancer on magnetic resonance imaging using a convolutional neural network algorithm. BJU international, 2018. 122(3): p. 411–417.
Yoo, S., et al., Prostate cancer detection using deep convolutional neural networks. Scientific reports, 2019. 9(1): p. 1–10.
Woźnicki, P., et al., Multiparametric MRI for prostate cancer characterization: Combined use of radiomics model with PI-RADS and clinical parameters. Cancers, 2020. 12(7): p. 1767.
Castillo T, J.M., et al., Automated classification of significant prostate cancer on MRI: a systematic review on the performance of machine learning applications. Cancers, 2020. 12(6): p. 1606.
Prostate MR Image Database, http://prostatemrimagedatabase.com/, Editor. accessed Aug. 22, 2020.
Mikołajczyk, A. and M. Grochowski. Data augmentation for improving deep learning in image classification problem. in 2018 international interdisciplinary PhD workshop (IIPhDW). 2018. IEEE.
Pawar, S. and V. Banga. Morphology Approach in Image Processing. in International Conference on Intelligent Computational Systems (ICICS’2012)(Dubai). Dubai. 2012.
Kovesi, P. Fast almost-gaussian filtering. in 2010 International Conference on Digital Image Computing: Techniques and Applications. 2010. IEEE.
Wong, C.Y., et al., Image contrast enhancement using histogram equalization with maximum intensity coverage. Journal of Modern Optics, 2016. 63(16): p. 1618–1629.
Anitha, S., et al. Analysis of filtering and novel technique for noise removal in MRI and CT images. in 2017 International Conference on Electrical, Electronics, Communication, Computer, and Optimization Techniques (ICEECCOT). 2017. IEEE.
Sreedhar, K. and B. Panlal, Enhancement of images using morphological transformation. arXiv preprint arXiv:1203.2514, 2012.
Huang, H., et al., Faster R-CNN for marine organisms detection and recognition using data augmentation. Neurocomputing, 2019. 337: p. 372–384.
Ren, S., et al., Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 2015. 28.
Girshick, R. and R. Fast, IEEE Int. Conf. Comput. Vis. Santiago, Chile, December, 2015: p. 7–13.
Redmon, J. and A. Farhadi. YOLO9000: better, faster, stronger. in Proceedings of the IEEE conference on computer vision and pattern recognition. 2017.
Liu, W., et al. Ssd: Single shot multibox detector. in European conference on computer vision. 2016. Springer.
Krizhevsky, A., I. Sutskever, and G.E. Hinton, Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 2012. 25.
Cho, J., et al., How much data is needed to train a medical image deep learning system to achieve necessary high accuracy? arXiv preprint arXiv:1511.06348, 2015.
Rácz, A., D. Bajusz, and K. Héberger, Effect of dataset size and train/test split ratios in QSAR/QSPR multiclass classification. Molecules, 2021. 26(4): p. 1111.
Alloghani, M., et al., A systematic review on supervised and unsupervised machine learning algorithms for data science. Supervised and unsupervised learning for data science, 2020: p. 3–21.
Yamashita, R., et al., Convolutional neural networks: an overview and application in radiology. Insights into imaging, 2018. 9(4): p. 611–629.
Szegedy, C., et al. Going deeper with convolutions. in Proceedings of the IEEE conference on computer vision and pattern recognition. 2015.
He, K., et al., Deep residual learning for image recognition (2015). cite. arXiv preprint arxiv:1512.03385.
Howard, A.G., et al., Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017.
Majnik, M. and Z. Bosnić, ROC analysis of classifiers in machine learning: A survey. Intelligent data analysis, 2013. 17(3): p. 531–558.
Powers, D.M., Evaluation: from precision, recall and F-measure to ROC, informedness, markedness and correlation. arXiv preprint arXiv:2010.16061, 2020.
Sokolova, M. and G. Lapalme. Classification of opinions with non-affective adverbs and adjectives. in Proceedings of the International Conference RANLP-2009. 2009.

No competing interests reported.

Classification of prostate cancer using Deep Learning approach and MobileNetV2 architecture

Status:

Version 1

Abstract

Figures

1. Introduction

2. Materials And Method

2.1 Data Preprocessing and Augmentation

2.2 Separating

2.3 Feature Extraction and Classification by Deep Learning Network

2.3.1 ResNet Neural Network

2.3.2 MobileNetv2 Neural Network

2.4 Performance Assessment Criteria

3. Results And Desiccation

4. Conclusion

Declarations

References

Additional Declarations

Status:

Version 1