Automated Diagnosing of Melanoma Based on the Stolz Algorithm

Abstract

The article discusses an algorithm for automating melanoma diagnosing using computer vision and a neural classifier. The aim of the study is to develop an algorithm for efficient classification of melanoma in images, as well as the influence of dermoscopic features on diagnosing. New growths in the image are highlighted. Calculation of parameters for the Stolz formula and calculation of TDS (calculation of diagnosis) were carried out. Based on the calculations performed, the optimization problem was solved to clarify the boundaries of the neoplasm differentiation areas, and the Stolz formula coefficients adjustment, and a set for neural network training was generated; MLP was selected as a classifier. The average accuracy of the algorithm on two independent data sets is 86%. Classification by a neural network of two sets of images according to Stolz parameters is given. Based on the results, a diagnosing was made. Based on the Stolz algorithm, a set of images with skin neoplasms is pre-processed: the neoplasm is isolated, ABCD signs and the overall dermoscopic score are calculated from the image. The proposed algorithm for automated melanoma diagnosing based on the Stolz dermoscopic method showed good results on real sets of images. The reduced accuracy of the model on the HAM10000 set is due to the class imbalance (ratio of classes 1 to 9). A high F-measure value indicated a good balance between sensitivity and specificity of the model, with the model tending to be better at detecting melanoma true positive (TP) cases. High sensitivity is desirable for problems, where reducing false negatives (missing positive cases) is critical, even at the expense of increasing Type I errors. The completeness of the model was 88.54% and 95% and suggests that the model minimizes cases of missing diseases. It is concluded that the more accurately the ABCD parameters are calculated, the more accurately the model will be able to classify the disease; in the future, refining and dividing parameter D into several components can provide the model with additional features for classification.