Artificial intelligence–assisted transrectal ultrasound for rectal cancer invasion depth classification

Abstract Background Accurate assessment of tumor invasion depth is essential for treatment selection in rectal cancer. Transrectal ultrasound (TRUS) provides high-resolution visualization of rectal wall layers but is highly operator-dependent, limiting reproducibility outside expert centers. Artificial intelligence–based approaches may help standardize TRUS interpretation, particularly for depth-of-invasion assessment. Methods We developed and evaluated a two-stage deep learning model for automated analysis of standard TRUS images. A total of 677 two-dimensional images from 103 patients were included. In Stage 1, images were classified as normal or cancer. Images identified as cancer were then analyzed in Stage 2 to classify invasion depth as superficial (T0–T1) or deep (T2–T4). Models were trained using transfer learning with five-fold cross-validation and evaluated on an independent held-out test set. Performance was assessed using accuracy, area under the receiver operating characteristic curve (AUC), and class-specific recall. Results On the independent test set, the two-stage pipeline achieved an overall three-class accuracy of 82.4%. Stage 1 discriminated normal from cancer with an accuracy of 89.7% and an AUC of 0.93, ensuring reliable identification of tumor-containing images. Among cancer-positive images, Stage 2 differentiated superficial from deep invasion with an accuracy of 85.6% and an AUC of 0.91. Performance was consistent across the dataset and comparable to reported accuracies of expert-performed TRUS. Conclusions A two-stage deep learning approach applied to standard TRUS images enables automated, practice-oriented classification of rectal tumor invasion depth. By reflecting real-world diagnostic workflows and reducing operator dependence, this method may support more consistent TRUS-based staging and assist treatment decision-making, particularly in nonspecialized centers.