A Method and Algorithm for Training a Convolutional Neural Network Designed for an Intelligent Melanoma Recognition System

The purpose of research is to develop an intelligent information system that allows to implement the process of melanoma diagnosis based on a convolutional neural network.

Methods. The basis for the formation of an intelligent diagnostic information system is the use of convolutional neural networks based on binary classification. During the study, a set of 21,000 images was analyzed, where 10,500 images were an image of a melanoma (malignant formation), and 10,500 images of a nevus (benign formation).As a result of the research, a module containing a convolutional neural network has been developed, which accepts informative signs of the analyzed image as input and forms a response consisting of a set of probabilistic characteristics of the selected pathological formation belonging to one of the classes of formations - melanoma or nevus. To ensure an increase in the level of classification accuracy, the convolutional neural network Xceptionwas used. The last layers of the neural network used were retrained, the network was regularized and the method of excluding neurons was implemented in order to reduce the loss function.

Results. Based on the processing of pathological objects in the images, a set of input information features designed for an intelligent recognition system has been formed. The use of convolutional neural network allowed to establish the accuracy of classification of pathological objects - 94.59%. The maximum value of the loss function during the research reached 18.79%. A module was formed representing a Linux container with an interface for interaction via the HTTP protocol. Based on the analysis of the input image, the module generates a response consisting of a probabilistic characteristic of the belonging of the studied object to one of the studied classes - melanoma and nevus.

Conclusion. The results of the sequential formation of a convolutional neural network by analyzing changes in key indicators (loss function and model accuracy) during the change in the size of the network layers are presented.

1. Dannye po onkologicheskim zabolevaniyam v Kurskoi oblasti [Data on oncological diseases in the Kursk region]. Available at: https://adm.rkursk.ru/index.php?id=13&mat_id=108524. (accessed 07.12.2021)

2. O Strategii tsifrovoi transformatsii klyuchevykh otraslei ekonomiki, sotsial'noi sfery i gosudarstvennogo upravleniya Kurskoi oblasti na period s 2021 po 2024 gody [On the Strategy of digital Transformation of key Sectors of the economy, social sphere and Public Administration of the Kursk region for the period from 2021 to 2024]. Resolution of the Administration of the Kursk region of August 08, 2021 N 880-pa. Available at: https://adm.rkursk.ru/in-dex.php?id=109&mat_id=123917. (accessed 22.12.2021)

3. Melanoma kozhi [Melanoma of the skin]. Available at: https://nmicr.ru/meditsina/onkologicheskie-zabolevaniya-i-programmy-lecheniya-raka/programma-protiv-raka-kozhi-i-skeleta/melanoma-kozhi. (accessed 11.01.2021)

4. Paukov V. S. Patologicheskaya anatomiya [Pathological anatomy]. Moscow, GEOTAR-Media Publ., 2015, vol. 2, pp. 359-360.

5. Stadii melanomy [Stages of melanoma]. Available at: https://daily-med.com.ua/stati/stati-chto-takoe-melanoma-kozhi-i-kak-ee-lechit. (accessed 12.01.2021)

6. Metod ABCDE v diagnostike melanomy [The ABCDE method in the diagnosis of melanoma]. Available at: https://melanomaunit.moscow/onkologiya/melanoma. (accessed 12.12.2021)

7. Klassifikatsiya melanomy po A. Breslou melanomy [Classification of melanoma according to A. Breslow melanoma]. Available at: https://melanomaunit.ru/melanoma-diagnostics/klassifikatsiya-melanomy. (accessed 13.12.2021)

8. Ivanov A. V., Tomakova R. A., Mishustin V. N. Intellektual'naya sistema morfologicheskogo analiza patologicheskikh izmenenii pri retinopatii [Intelligent system of morphological analysis of pathological changes in retinopathy]. Izvestiya Yugo-Zapadnogo universiteta. Seriya: Upravlenie, vychislitel'naya tekhnika, informatika. Meditsinskoe priborostroenie = Proceedings of the Southwest State University. Series: Control, Computer Engineering, Information Science. Medical Instruments Engineering, 2012, no. 2-1(41), pp. 19-27.

9. Tomakova R. A., Filist S. A., Tomakov M. V. Gibridnye tekhnologii v intellektual'nykh sistemakh identifikatsii lekarstvennykh sredstv [Hybrid technologies in intelligent drug identification systems]. Nejrokomp'yutery: razrabotka, primenenie = Neurocomputers: Development, Application, 2014, no. 6, pp. 31-66.

10. Tomakova R. A., Filist S. A., Pykhtin A. I. Development and Research of Methods and Algorithms for Intelligent Systems for Complex Structured Images Classification. Journal of Engineering and Applied Sciences, 2017, vol. 12, no. 22, рр. 6039-6041.

11. Korenevsky N. A., Tomakova R. A., Seregin S. P., Rybochkin A. F. Nejronnyeseti s makrosloyami dlya klassifikaciii prognozirovaniya patologij setchatki glaza [Neural networks with macro layers for classification and prediction of retinal pathologies]. Medicinskaya tekhnika = Medical Equipment, 2013, no. 4(280), pp. 16-18.

12. Tomakova R., Filist S., Veynberg R., Brezhnev A., eds. The Role of Hybrid Classifiers in Problems of Chest Roentgenogram Classification. Advance in Intelligent Systems and Computing, 2020, vol. 902, pp. 293-303.

13. Tomakova R. A., Filist S. A., Pykhtin A. I., Shutkin A. N. Intelligent Medical Decision Suport System Based on Internet-Technology // 16th International Multidisciplinary Scientific Geoconference SGE 2016. Albena, 2016, рр. 263-270.

14. Tomakova R. A., Efremov M. A., Zhilin V. V. Echetkie neirosetevye tekhnologii dlya klassifikatsii formennykh elementov krovi [Fuzzy neural network technologies for classification of shaped blood elements]. Izvestiya Yugo-Zapadnogouniversiteta. Seriya: Upravlenie, vychislitel'naya tekhnika, informatika. Meditsinskoe priborostroenie = Proceedings of the Southwest State University. Series: Control, Computer Engineering, Information Science. Medical Instruments Engineering, 2013, no. 4, pp. 84-89.

15. The HAM10000 Dataset, Alarge Collection of Multi-Sourcedermatos Copicimages of Common Pigmented Kinlesions. Available at: https://www.nature.com/articles/sdata2018161. (accessed 14.12.2021)

16. Melanoma data set 256x256px. Available at: https://www.kaggle.com/cdeotte/jpeg-melanoma-256x256/. (accessed 13.12.2021)

17. Keras Applications. Available at: https://keras.io/api/applications/. (accessed 14.01.2021)

18. Adam - latest trends in deep learning optimization. Available at: https://towardsdatasci-ence.com/adam-latest-trends-in-deep-learning-optimization-6be9a291375c. (accessed 14.12.2021)

19. Dropout. Available at: https://keras.io/api/layers/regularization_layers/dropout/. (accessed 14.12.2021)

20. Layer weight regularizers. Available at: https://keras.io/api/layers/regularizers/. (accessed 14.12.2021)