Dual modal pathomics model for colorectal cancer early recurrence prediction and mutation landscape analysis.

Patients with stage II/III colorectal cancer remain at substantial risk of early postoperative recurrence, yet accurate risk stratification remains challenging. Here, we developed a dual-modal pathomics model integrating hematoxylin-eosin and Ki-67 whole-slide images to predict early recurrence in 362 patients from two medical centers. Among multiple pretrained feature encoders, the integrated hematoxylin-eosin plus Ki-67 model using the UNI encoder achieved the best performance, with an area under the curve of 0.902 in the external validation cohort. The model consistently stratified patients into distinct prognostic groups across clinical subgroups. Attention map visualization further suggested that high-risk predictions were associated with tumor invasive fronts, whereas low-risk predictions were linked to immune infiltration and fibrotic stromal regions. These findings highlight the potential of multimodal pathology artificial intelligence for clinically interpretable prognostic assessment and personalized postoperative management in colorectal cancer.