Deep learning in enhanced CT imaging: predicting invasion depth of rectal adenocarcinoma

This study aims to develop and evaluate a deep learning model, RectoDepthAI, that leverages enhanced CT images to accurately assess the tumor invasion depth in rectal adenocarcinoma, distinguishing between early-stage (submucosal or muscularis propria invasion) and advanced-stage (pararectal tissue or adjacent structure invasion) tumors.

Utilizing a dataset of 934 patients, RectoDepthAI integrates ResNet-18 for spatial feature extraction and Long Short-Term Memory (LSTM) for sequential processing of venous phase CT slices. Data management involved preprocessing, augmentation, and 5-fold cross-validation; evaluation metrics included AUC, accuracy, sensitivity, and specificity; interpretability was enhanced via Gradient-weighted Class Activation Mapping (Grad-CAM).

In its evaluation phase, RectoDepthAI was assessed over five rounds of training and validation, demonstrating robust performance with an average AUC of 0.883, accuracy of 84.1%, sensitivity of 87.0%, and specificity of 80.0%. Grad-CAM enhanced the interpretability of diagnostic predictions.

RectoDepthAI emerges as a promising tool for non-invasive staging of rectal adenocarcinoma, improving diagnostic precision and supporting clinicians in treatment decision-making.