Pre-treatment T2WI-based automated artificial intelligence system for predicting neoadjuvant chemoradiotherapy response and prognosis in locally advanced rectal cancer: a multicenter study

In clinical practice, approximately 15–25% of patients with locally advanced rectal cancer (LARC) achieve a pathological complete response (pCR), defined as tumor regression grade 0 according to the AJCC 8th edition, which is closely associated with favorable prognosis. However, noninvasive and accurate preoperative assessment methods remain a major clinical challenge. We aimed to develop an automated artificial intelligence (AI) system for predicting the response to neoadjuvant chemoradiotherapy and prognosis in LARC using pre-treatment T2WI. This multicenter retrospective study included 227 LARC patients from Center 1 as the development cohort and 72 and 55 patients from two separate centers as external validation cohorts 1 and 2, respectively. 3D Swin-UNETR was constructed to automatically identify LARC tumors. Based on the segmented tumor volumes, three analytical approaches were implemented to support downstream tasks. Specifically, 1,132 radiomic features (Rf) were extracted using PyRadiomics, while 2,048 deep learning features (Dlf) were derived by 3D ResNet-152 architecture. Following a series of feature selection steps, predictive response scores were generated, including RF scores (Rfs) and Dlf scores (Dlfs). Additionally, an end-to-end analysis was performed to automatically generate Dl scores (Dls). A multi-perspective feature engineering model was then developed by integrating Rfs, Dlfs, and Dls. The predictive performance was evaluated using the area under the receiver operating characteristic curves (AUROCs). Prognostic indices, including disease-free survival (DFS) and overall survival (OS), were assessed using Kaplan-Meier (KM) analysis, with the Log-Rank test used to compare KM curves. The fusion model achieved moderate AUROCs of 0.756 and 0.751 in external validation cohorts 1 and 2, respectively. It effectively stratified OS risk, with p-values < 0.05 across all cohorts. For DFS, statistical significance was inconsistent across cohorts, being achieved in the development cohort ( p = 0.0084) and external cohort 2 ( p = 0.044), but not in external cohort 1 ( p = 0.068). These findings are exploratory and require prospective confirmation. The proposed AI system demonstrates moderate, externally validated performance for assessing treatment response and OS risk stratification in LARC and may serve as a decision-support tool to assist clinical judgment.