Stages of training neural networks for classification and detection of skin neoplasms

BACKGROUND: In recent years, neural networks have become an integral part of many fields, including medicine. However, the effectiveness of these models directly depends on the quality of the training data on which they are trained. Creating and maintaining a high-quality training dataset is a critical step in the development process of neural networks. AIM: The aim of the research is to identify the key characteristics of the training database for the neural network that influence its subsequent sensitivity and specificity. MATERIALS AND METHODS: A database of verified images of skin neoplasms was created to train a neural network to implement it in large-scale screening examinations. In the first phase of the study, a database was created to train a neural network to classify images of skin neoplasms (NSCa). Between 2017 and 2019, 7,680 digital images were collected from 6,892 patients with verified diagnoses: 5,316 (69,22%) confirmed by pathological examination, and 2,364 (30,78%)) confirmed clinically and dermatoscopically. A dataset containing 7,680 verified clinical images of skin neoplasms was created, and 1,680 images constituted the test sample for analyzing the model's effectiveness. The performance indicators of NSCa were as follows: sensitivity (Se): 70.47%; specificity (Sp): 79.86%; diagnostic accuracy (Ac): 74.68%. Due to the low sensitivity and specificity rates, the following steps were taken: (1) an additional round of training was conducted; (2) image quality control methods were developed; (3) a detection neural network was created, and (4) a new neural (NSCb) was established. RESULTS: The neural network, trained on a verified dataset of clinical images of benign and malignant skin neoplasms and having undergone multiple rounds of training, operates with a sensitivity of 85.32–86.97% and a specificity of 87.59–88.92%. These rates exceed the sensitivity and specificity of skin neoplasm diagnoses made by non-oncological specialists using the naked eye, allowing for the use of this method in population screening. Following the retraining of the neural network and the establishment of NSCb, the creation of neural network, and the development of image quality control methods, an increase in the sensitivity and specificity of the neural network's performance was observed. CONCLUSION: The use of artificial intelligence as a physician's assistant imposes quite high requirements on the performance parameters of the neural network. Mechanical learning, even on a large volume of verified data, did not achieve the desired results. The sequential work aimed at improving the parameters involved conducting an additional round of training, developing image quality control methods, and creating a detection neural network and a classification neural network. As a result, the trained neural network operates with a sensitivity of 85.32% to 86.97% and a specificity of 87.59% to 88.92%, which has enabled the use of the trained neural network as a tool for population screening.