Integrative Multimodal Deep Learning for Individualized Recurrence Risk Stratification in Stage I–III Colon Cancer

Melissa de Bruin, Myriam Chalabi, Xinglong Liang, Milan Pijl, Geerard Beets, Tianyu Zhang, Regina Beets-Tan

Keywords: Multimodal Deep Learning, Colon Cancer Recurrence, Contrastive Pre-training, Attention Mechanism, Prognostic Stratification

TL;DR: A multimodal deep learning model integrating CT imaging and clinical data improves recurrence risk stratification in stage I–III colon cancer.

Registration Requirement: Yes

Abstract: Precise risk stratification in stage I–III colon cancer remains a clinical challenge, as conventional radiological staging often fails to identify high-risk patients. This study developed a multimodal deep learning model integrating preoperative CT imaging with clinical data to predict recurrence. In a retrospective cohort of 713 patients from two centers, the model utilized a convolutional neural network with an attention mechanism for image feature extraction and a multilayer perceptron for clinical variable processing. A late fusion strategy was employed to generate a unified risk score. The model achieved a C-index of 0.665 and a 2-year recurrence AUC of 0.766. Stratification by median risk score yielded a significant hazard ratio of 2.46. These findings suggest that integrating heterogeneous data sources through deep learning provides a robust prognostic tool, potentially facilitating personalized therapeutic interventions.

Visa & Travel: No

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 128