Deep Learning Assisted Cell Morphology Analysis for Automated Disease Screening

Abstract

Microscopic blood analysis and tissue analysis Enhanced with cell morphology Automated analysis Cell morphology analysis is a vital part of contemporary diagnostic medicine. In this paper, I have introduced MorphoViT, a new deep learning architecture that combines Vision Transformer (ViT) encoders with a multi-scale convolutional segmentation head, as a structure that can be used to analyze the morphology of the cell and perform disease screening with high accuracy. The strategy presented has three important innovations, namely: (i) a Morphology-Aware Attention (MAA) module with dynamic focus on discriminative sub-cell structures; (ii) contrastive pre-training with unlabeled peripheral blood smear images to improve the feature representations; and (iii) a stain-invariant normalization network that guarantees the cross-laboratory generalizability. MorphoViT has the highest classification accuracy of 97.8 %and segmentation Dice score of 0.934 and F1-score of 0.961, making it the best existing state-of-the-art approach by up to 3.7% in terms of three publicly available benchmarks: BloodMNIST, the Matek-19 leukocyte dataset, and the ISBI Cell Tracking dataset