Preoperative CT deep learning of tumor and visceral adipose tissue combined with postoperative pathology predicts colorectal cancer recurrence: a multicenter study.
Authors
Affiliations (3)
Affiliations (3)
- Department of Radiology, Jiangxi Cancer Hospital & Institute, Jiangxi Clinical Research Center for Cancer, The Second Affiliated Hospital of Nanchang Medical College, Nanchang, China.
- Department of Radiology, The First Affiliated Hospital, Jiangxi Medical College, Nanchang University, Nanchang, China.
- Department of Radiology, The Second Affiliated Hospital, Jiangxi Medical College, Nanchang University, Nanchang, China.
Abstract
To develop and externally validate a preoperative CT imaging model integrating primary tumour and visceral adipose tissue (VAT) features for predicting recurrence in colorectal cancer (CRC). This retrospective multicentre study included 1,369 patients with CRC undergoing upfront curative-intent resection, comprising a training cohort (n = 705) and two external test cohorts (n = 425 and 239). A 3D convolutional neural network extracted tumour-centred volumetric features, while an attention-based multiple instance learning branch aggregated whole-abdominal VAT slices into patient-level representations. The two CT-derived scores constituted the preoperative imaging model. A clinicopathological component was combined with the imaging scores to construct the postoperative multimodal model; robustness to variable selection was evaluated using an all-candidate ridge-penalised Cox sensitivity model. Both tumour and VAT DL scores remained independently associated with recurrence-free survival after multivariable adjustment. In external test cohorts 1 and 2, the preoperative Tumor + VAT imaging model achieved C-indices of 0.734 and 0.772, whereas the postoperative multimodal model achieved C-indices of 0.772 and 0.810 and 2-year time-dependent AUCs of 0.757 and 0.836, respectively. Training-derived tertile cut-offs from the postoperative model consistently stratified patients into low-, intermediate- and high-risk groups across all cohorts (all log-rank P < 0.001). In the two external cohorts, the clinicopathological-only model achieved C-indices of 0.709 and 0.715, compared with 0.772 and 0.810 for the postoperative multimodal model. The preoperative CT model non-invasively stratified recurrence risk using tumour and whole-abdominal VAT phenotypes, while postoperative clinicopathological features further improved prognostic performance.