The potential of an artificial intelligence for diagnosing MRI images in rectal cancer: multicenter collaborative trial.

Background: An artificial intelligence-based algorithm we developed, mrAI, satisfactorily segmented the rectal tumor, rectum, and mesorectum from MRI data of rectal cancer patients in an initial study. Herein, we aimed to validate mrAI using an independent dataset.

Methods: We utilized MRI images collected in another nationwide research project, "Open versus Laparoscopic Surgery for Advanced Low Rectal Cancer Patients". MRIs from 467 cases with upfront surgery were utilized; six radiologists centralized the MRI evaluations. The diagnostic accuracies of mrAI and the radiologists for tumor depth were compared using pathologic diagnosis as a reference.

Results: For all cases, centralized diagnosis demonstrated 84.2% sensitivity, 37.7% specificity, and 73.7% accuracy; mrAI exhibited 70.6% sensitivity, 61.3% specificity, and 68.5% accuracy. After limiting MRIs to those acquired by a Philips scanner, with an inter-slice spacing of ≤ 6 mm-both conditions similar to those used in the development of mrAI-the performance of mrAI improved to 76.8% sensitivity, 76.7% specificity, and 76.7% accuracy, while the centralized diagnosis showed 81.8% sensitivity, 36.7% specificity, and 71.3% accuracy. Regarding relapse-free survival, the prognosis for tumors staged ≥ T3 was significantly worse than for tumors staged ≤ T2 (P = 0.0484) in the pathologic diagnosis. While no significant difference was observed between ≥ T3 and ≤ T2 tumors in the centralized diagnosis (P = 0.1510), the prognosis for ≥ T3 was significantly worse in the mrAI diagnosis (P = 0.0318).

Conclusion: Proper imaging conditions for MRI can enhance the accuracy of mrAI, which has the potential to provide feedback to radiologists without overestimating tumor stage.