Sergey A. Milantev, Vitaliia I. Sviatkina, Igor A. Bessmertny, Kirill V. Zaichenko

ITMO University, IAI RAS

This research explores the possibility of applying combined convolutional neural network architectures to analyze skin lesions. Model architectures have been developed to extract additional features related to the shape pattern of skin lesions. Optimization of the models, including the architecture, was performed in order to minimize I and II types of errors for rare skin lesions. ISIC2017-2020, MED-NODE, SD-198, 7-Point Criteria Database, Light Field Image Dataset of Skin Lesions, PH2, IAP RAS were used in the training process. AdamW optimizer, FocalLoss functions and CosineAnnealingWarmRestarts scheduler were used to train classification models. The BCEDice loss function was used to train the segmentation models. The models were evaluated using weighted classification metrics such as Specificity, Recall, Precision and F1-score. The robustness of model architecture was considered during the validation phase. Models which are using additional convolutional neural networks for the skin lesion extraction shape features showed better metrics performance and also had lower sum of I and II type errors for rare lesions compared to conventional classification models. The results of this research can be used in analyzing medical problems with data imbalance in the training dataset.

skin lesions, convolutional neural networks, skin lesion analysis, class imbalance, multispectral image processing