Deep Learning-Based Segmentation and Paracancerous Tissue Analysis in Colorectal Cancer Histopathology

Colorectal cancer (CRC) is a globally prevalent and lethal malignancies worldwide. Early detection plays critical role in improving therapeutic outcomes. However, there remains a significant lack of specialized, directly applicable, and publicly available datasets specifically tailored for CRC histopathological image segmentation. Despite advancements in computer vision and the demonstrated success of deep learning, it remains an open question how effectively these technologies can be leveraged to improve the performance of CRC image segmentation. To bridge this gap, we developed and released a publicly accessible dataset comprising endoscopic biopsy Hematoxylin and Eosin (H&E) stained histopathological images, named EBHI-Seg. This dataset supports both the training and evaluation of image segmentation models. We benchmark segmentation performance on EBHI-Seg using both classical machine learning approaches and state-of-the-art deep learning techniques. Furthermore, we propose an enhanced Trans-UNet architecture designed to improve the recognition of fine-grained features in complex pathological images. To facilitate practical application and support early CRC diagnosis, we also developed a user-friendly graphical user interface (GUI) that integrates the proposed deep learning models. The system enables automated analysis of pathological images and quantification of clinically relevant parameters such as intra-tumoral TILs (i-TILs) and stromal TILs (s-TILs).