Skin lesion classification using fuzzy-based ensemble and decentralised federated learning approaches

Abstract

Early treatment is crucial for recovery from skin cancer. Detection of malignant skin lesions is required for timely diagnosis of skin cancer and to reduce mortality from this disease. This project explores convolutional neural networks(CNN) to classify skin lesions by leveraging Sugeno Integral for aggregation of CNN classifier prediction scores and self attention mechanism for improving base model performance. The use of Sugeno Integral allows for effective aggregation of multiple CNN models, capturing diverse features and improving classification accuracy. Self attention assigns weighted importance to each region of the image. This allows the model to focus on critical areas of the image leading to better capture of critical patterns. The attention mechanism is integrated with the Sugeno ensemble for improved accuracy. Due to privacy concerns, skin cancer data cannot be moved to a central location for model training. In addition to the centralized model training, we introduce a collaborative peer-to-peer decentralised federated learning approach to model training without using a central server. This method transfers the model updates between clients while the data remains stored locally. The proposed models were tested on the ISIC 2016 and the HAM10000 datasets. Ensemble of prediction scores provided better results compared to the base models for both datasets. The enesemble of models with self attention applied provided the best results due to attention improving base performances. The federated learning model was the least performing but still comparable to the ensembles making it an approach worth considering for distributed data. Keywords: Skin Cancer, Deep Learning, Convolutional Neural Network, Fuzzy measures, Sugeno Integral, Ensemble, Self Attention, Decentralized, Federated Learning