Deep learning-enabled multiphoton microscopy predicts colorectal cancer recurrence from routine FFPE specimens.

Publication Information

Title: Deep learning-enabled multiphoton microscopy predicts colorectal cancer recurrence from routine FFPE specimens.

Author(s):Yang Y, Xiao C, Zou D, Wang L, Yang R, Zhang Y, Zhang L, Zhao Z, Qiu S, Liu S, Bai Y, Sun WY, He RR, Chen G, Li T, Luo OJ, Jiang W.

‌Journal Name, Year, Volume(Issue): Page range: NPJ Digit Med. 2025 Nov 18;8(1):689

DOI:10.1038/s41746-025-02058-3

Abstract:Colorectal cancer recurrence remains a major challenge after curative resection, and accurate tools for early risk assessment are essential to stratify patients and guide personalized therapeutic planning. We developed MPMRecNet, a dual-stream deep learning model for predicting recurrence using multiphoton microscopy imaging of formalin-fixed paraffin-embedded tissue sections from 1071 patients across two hospitals. MPMRecNet employs MaxViT-based encoders, cross-modal attention fusion, and classification under focal loss with mixed-precision optimization. It achieved strong external validation performance (ROC-AUC = 0.849, PR-AUC = 0.664), outperforming traditional clinical predictors. Multivariable analysis confirmed MPMRecNet as the most powerful independent predictor of recurrence (OR = 5.66, p < 0.001), and a combined nomogram incorporating clinical variables further improved stratification (ROC-AUC = 0.872). MPMRecNet offers a non-destructive tool for recurrence prediction from routine pathology slides, supporting precise risk assessment and postoperative surveillance.

First Page of Article: