STAR: Spatio-Temporal Attention-guided Recurrence Prediction for Colorectal Cancer Liver Metastasis

LIZESHENG, Wei Yang, Yan Su, Yiran Zhu, Yuhan Tang, Haoran Chen, Pan Chengchang, Honggang Qi

Keywords: Colorectal cancer liver metastasis, Dynamic risk prediction, Postoperative management, Deep learning framework

TL;DR: We developed a deep learning model called STAR that analyzes a patient's CT scans over time to dynamically predict when and where their colorectal cancer is most likely to return each year after surgery.

Abstract: Postoperative recurrence remains a primary obstacle to long-term survival for patients with colorectal cancer liver metastasis (CRLM).However,a major limitation of current prognostic methods is their static nature, which fails to account for the dynamic, time-evolving risk of recurrence driven by complex postoperative processes like liver regeneration and microenvironmental shifts. To overcome this challenge, we therefore developed STAR, a novel deep learning framework designed for dynamic prediction of postoperative recurrence and survival. Specifically, the model's primary innovation is its ability to forecast long-term, year-by-year prognosis by analyzing the temporal evolution of postoperative CT scans in conjunction with key clinical data. Additionally, the framework can simultaneously generate personalized, annual recurrence risk heatmaps. These heatmaps offer an intuitive, visual guide to the probable location and timing of recurrence, thereby providing clinicians with interpretable, patient-specific insights to tailor dynamic surveillance strategies. When evaluated on the MSKCC CRLM dataset, our model demonstrated outstanding predictive performance, achieving 90% accuracy in assessing survival status for each of the 12 postoperative years (Temporal Adjacency Accuracy, TAA) with a Mean Absolute Error (MAE) of 0.7500. This study consequently establishes a new paradigm for the postoperative management of CRLM, shifting the focus from static, single-point assessment to continuous, dynamic risk monitoring. Ultimately, by providing a tool for more precise and personalized follow-up, our framework helps advance the clinical goal from simple "spatial resection" to a more comprehensive "spatiotemporal cure."

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 11502