The pronounced heterogeneity of the tumor microenvironment (TME) in colorectal cancer (CRC) presents major obstacles in accurately predicting patient outcomes and tailoring treatment responses. Deciphering this intricate microenvironment based on histological images and classifying it into well-defined tissue components is critical for optimizing clinical interventions. Although deep learning (DL) has advanced substantially in medical imaging analysis, its application in CRC remains limited due to a shortage of comprehensively annotated datasets and large-scale, high-quality histological images. To address this gap, we present HMU-CRC-Hist550K, a curated dataset comprising 550,000 annotated image tiles derived from 500 whole-slide images, fully labeled into eight distinct TME tissue classes. The dataset represents a broad collection of publicly available CRC histology samples. Additionally, we demonstrate the utility of this resource by benchmarking three DL models on tissue segmentation tasks. HMU-CRC-Hist550K offers a valuable foundation for TME profiling, AI-assisted diagnosis, molecular subtype inference, and individualized therapy planning, while also enabling new research directions in modeling the spatial-temporal evolution of the TME.
Wang et al. (Thu,) studied this question.