Background: Loco-regional staging of rectal cancer relies on MRI and rectal EUS (R-EUS). In situ and T1 tumors may be candidates for endoscopic resection, and R-EUS enables reliable differentiation among early-stage tumors (Tis, T1, T2), a distinction that MRI cannot consistently provide. This prospective pilot study aimed to evaluate the predictive performance of a deep learning (DL) tool for detecting and staging rectal tumors on R-EUS. Methods: The DL tool uses a convolutional neural network for image segmentation and classification. Performance in lesion segmentation was assessed using the Dice Similarity Coefficient (DSC) and F1-score. The model was first evaluated for its ability to differentiate in situ and T1 tumors from T2/T3 lesions, and then for distinguishing Tis from other stages (T1, T2, T3). Results: Fifty patients were enrolled. The median DSC for tumor segmentation was 0.65 (IQR 0.17). Tumor detection showed an accuracy of 0.77, precision of 0.85, recall of 0.77, and F1-score of 0.81. In distinguishing Tis/T1 from T2/T3 tumors, the model achieved an accuracy of 0.64, precision of 0.88, recall of 0.64, and F1-score of 0.74. For distinguishing Tis from T1, T2, and T3 lesions, the accuracy was 0.80, precision 0.83, recall 0.89, and F1-score 0.86. Mesorectal lymph node segmentation showed a median DSC of 0.62 (IQR 0.17). Conclusions: The DL tool shows promise for aiding operators in identifying rectal lesions suitable for endoscopic resection. The semi-supervised training approach reduces manual segmentation burden and achieves performance comparable to expert physicians.
Montale et al. (Tue,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: