Decentralized medical image classification system using dual-input CNN enhanced by spatial attention and heuristic support

The Internet of Medical Things (IoMT) enables the construction of expert systems that can use shared resources or data. In this work, we present a new IoMT operation model based on decentralized federated learning, which has been extended with a blockchain module. To ensure the security of medical data, we are introducing local and global blockchains with committee consensus mechanisms. Moreover, for data classification, we propose the use of a dual -input network with a spatial attention module which allows for feature fusion of processed images. The processed images are the original sample and one modified by a heuristic algorithm to extract global features in the form of superpixels. The use of this network model allows for the extraction of other features from the image in separate branches, which are processed by the attention module and subsequent fusion. The proposed system model solution with a newly modeled network was analyzed and tested on publicly available databases to verify its operation. Based on the conducted experiments, the effectiveness of the proposal in comparison to state -of -art is significant and the classifier reached 0.9862 of accuracy using a decentralized proposed approach. Moreover, the region of superpixel can be a valuable asset to image processing tools.