Automated deep learning model for predicting pathological complete response in rectal cancer: A tool to organ-preserving strategies.
Authors
Affiliations (10)
Affiliations (10)
- University of Valencia, Valencia, Spain. [email protected].
- Departament of General and Digestive Surgery, Department of Colorectal Surgery, Clinic Universitary Hospital of Valencia, Av. Blasco Ibáñez, 17. 46010, Valencia, Spain. [email protected].
- INCLIVA Biomedical Research Institute, Valencia, Spain. [email protected].
- University of Valencia, Valencia, Spain.
- Departament of General and Digestive Surgery, Department of Colorectal Surgery, Clinic Universitary Hospital of Valencia, Av. Blasco Ibáñez, 17. 46010, Valencia, Spain.
- INCLIVA Biomedical Research Institute, Valencia, Spain.
- Department of Medical Oncology, INCLIVA Biomedical Research Institute, University of Valencia, Valencia, Spain.
- CIBERONC, Carlos III Health Institute, Madrid, Spain.
- Division of Hematology/Oncology, Mass General Brigham Cancer Center, Harvard Medical School, Boston, MA, USA.
- Radiology Department, Clinic Universitary Hospital, Valencia, Spain.
Abstract
Accurate identification of pathological complete response (pCR) after neoadjuvant chemoradiotherapy (nCRT) in locally advanced rectal cancer (LARC) remains a key clinical challenge. Clinical complete response is an imperfect surrogate, and pCR can only be definitively established after surgery. We developed a fully automated, segmentation-free deep learning model to support post-treatment response assessment using routine T2-weighted MRI. A longitudinal three-dimensional (3D) siamese convolutional neural network was trained using paired pre- and post-nCRT axial T2-weighted MRI volumes and a normalized signed voxel-wise difference map. The multitask framework simultaneously predicted rectal wall response (good response: modified Ryan score 0-1 vs poor response ≥ 2) and nodal status (ypN0 vs ypN +), from which pCR probability (ypT0N0) was derived. A retrospective single-center cohort of 195 patients was divided into training and independent test sets stratified by pCR status. Performance was evaluated using AUC-ROC and standard classification metrics with bootstrap-derived 95% confidence intervals. In the independent test set (n = 49; pCR prevalence 18.5%), the model achieved an AUC-ROC of 0.71 (95% CI: 0.55-0.85) for pCR prediction. At the selected operating threshold, sensitivity was 100% (95% CI: 70.1-100) and negative predictive value (NPV) was 100% (95% CI: 81.6-100), with a specificity of 42.5% (95% CI: 28.5-57.8). The high NPV reflects the low prevalence of pCR in the study cohort and may vary across external populations. This fully automated longitudinal deep learning model demonstrated moderate discrimination and a high-sensitivity profile for pCR detection. Its performance suggests potential utility as a screening or triage tool to support multidisciplinary assessment, rather than to directly guide organ-preserving strategies. External multicenter validation is required before clinical implementation.