Resident Vision Transformer: Lightweight Deep Learning Model for Disease Diagnosis on Edge Devices

Abstract

The deployment of AI-based medical image diagnosis on mobile edge devices face dual challenges of aliasing due to resizing images to lower resolutions, and huge total trainable parameters associated with deep learning architectures. This paper aims to address these challenges by proposing Resident Vision Transformer architecture with residual and dense connections enabling enhanced performance with lower number of parameters. The input to the proposed architecture is processed with adaptive padding, which not only maintains a constant input size, but also, preserves the spatial information. The proposed architecture has been trained and tested on three publicly available datasets, namely, breast cancer, skin cancer and brain tumor datasets. On the mini-MIAS database of mammograms, the proposed architecture achieved an accuracy of 92.94%, outperforming several related works. Similarly, on the PH2 database for Skin Cancer and the brain tumor dataset, the proposed architecture achieved an accuracy of 94.73% and 98.79%, respectively with fewer parameters. The proposed architecture paves way for feasibility of AI-driven medical image diagnosis on resource-constrained mobile edge devices. © 2024 IEEE.