A Deep Learning Model for Assessing Ki-67 and Tumor-Infiltrating Lymphocytes Using Multiscale Attention and Pixel Channel Fusion

Abstract

Nuclear protein Ki-67 and tumor-infiltrating lymphocytes (TILs) are key markers for histopathological diagnosis. They help to predict disease progression, patient prognosis, and treatment response. Recently, deep learning methods such as convolutional neural networks (CNNs) and transformers have significantly improved medical image segmentation tasks, making pathological diagnoses more efficient and accurate. However, many current methods fail to capture global information and use complex networks, leading to feature redundancy and high computational costs. To solve these problems, we propose a new U-Net-based method called Multi-Scale Attention and Pixel Channel Fusion Network (MAPC-Net). The encoder in MAPC-Net uses a Multi-Scale Attention Module (MA Module) to capture various cellular features by processing inputs at different scales and including both local and global spatial features. The decoder has a Pixel Channel Fusion Module (PCF Module) to combine channel and pixel-level information, reducing spatial detail loss and improving segmentation accuracy. Our tests on the SHIDC-BC-Ki-67 dataset show that MAPC-Net outperforms current top methods, with better F1 scores and lower root mean square error (RMSE). This advancement highlights MAPC-Net’s potential to improve the accuracy and reliability of histopathological image analysis, leading to better patient outcomes.

Keywords

• Breast cancer
• cell nuclei
• deep learning
• medical image segmentation