Enhanced digital pathology image recognition via multi-attention mechanisms: the MACC-Net approach

Abstract

Abstract Digital pathology has revolutionized cancer diagnosis through microscopic analysis, yet manual interpretation remains hindered by inefficiency and subjectivity. Existing deep models for osteosarcoma cell nucleus recognition suffer from the difficulty of capturing hierarchical relationships in single-dimensional attention mechanisms, leading to inaccurate edge recognition. Furthermore, the fixed receptive field of CNNs limits the aggregation of multi-scale information, hindering the differentiation of overlapping cells. This study introduces MACC-Net, a novel multi-attention based method designed to enhance the recognition accuracy of digital pathology images. By integrating channel, spatial, and pixel-level attention mechanisms, MACC-Net overcomes the limitations of traditional single-dimensional attention models, improving feature consistency and receptive field expansion. Experimental results demonstrate a Dice Similarity Coefficient (DSC) of 0.847, highlighting MACC-Net’s potential as a reliable auxiliary diagnostic tool for pathologists. Code: https://github.com/GFF1228/MACCNet.