Decentralized federated learning and deep learning leveraging XAI-based approach to classify colorectal cancer

Abstract

Convolutional Neural Networks based automated approaches are vastly utilised to anticipate and diagnose cancer, saving time and reducing mistakes. Deep Learning CNN methods use a variety of probabilistic and statistical methodologies to make computers understand and identify patterns in datasets based on previous experiences. We proposed federated learning(FL) based model to classify histopathological images for detecting colorectal cancer efficiently while providing high pre-diction accuracy. FL solves the problem of retaining privacy while utilizing vast and heterogeneous private datasets collected from numerous healthcare facilities. As the amount of patient data obtained is significantly responsible for the success of enhancing the accuracy of the system, the experiment was performed on a large dataset including cancerous and non-cancerous colorectal tissue images. FL is also capable of mitigating costs resulting from traditional ML approaches. Moreover, we have applied XAI method, a model-agnostic approach to acquire an explicit demonstration of the applied machine learning models. With XAI, we can visualize the super pixels of our colorectal tissue images through accepting and rejecting features. Applying vari-ous CNN models such as VGG, InceptionV3, ResNet, ResNeXt, and comparing their precision, we ascertained that ResNeXt50 bears the highest accuracy of 99.53%. Hence, we have applied ResNeXt50 on FL that brings forth the accuracy of 96.045% and F1 Score of 0.96.