Use of Artificial Intelligence in Analysis of Endoscopic Images Following Complete Clinical Response in Rectal Cancer

Background Patients with locally advanced rectal cancer (LARC) who have a complete clinical response (cCR) to neoadjuvant chemoradiotherapy (nCRT) may opt for organ preservation, wait and watch – (W&W). This consists of an intense surveillance program including serial endoscopies, pelvic MRI, CEA, and CT scans to detect recurrent disease at an early stage. However, identifying residual or recurrent lesions endoscopically in these cases can be challenging due to mucosal changes such as friability and neovascularization. We developed a novel deep learning model to assist in the detection of residual or recurrent rectal cancer lesions during proctosigmoidoscopy. Methods We trained a convolutional neural network (Wide ResNet-101-2) on a dataset of 1,795 annotated frames from proctosigmoidoscopy exams of 97 patients treated at a tertiary referral centre. Residual or recurrent disease was defined by histopathological confirmation. The dataset was split into training and testing cohorts using a 90/10% patient-level split. Results Out of 97 patients, 24 (363 frames) had confirmed residual or recurrent disease, while 73 (1,432 frames) presented normal rectal mucosa. The model achieved an overall accuracy of 92.8%, with a sensitivity of 80.0%, specificity of 97.3%, PPV of 90.9%, NPV of 94.0%, and an AUROC of 0.886. Conclusion To the best of our knowledge, this is the first deep learning model specifically developed for the detection of residual or recurrent disease following in W&W patients during endoscopic examination. This tool has the potential to enhance early lesion detection, guide clinical decision-making, and increase opportunities for salvage, curative treatment strategies.