The Diagnostic Value of Deep Learning for Multi-Classification of Rectal Cancer T Staging Based on Regional Attention

Objective: To explore the feasibility and effectiveness of an enhanced CT deep learning model based on regional attention for the preoperative multi-classification of rectal cancer T stages. Methods: Five hundred eligible patients with rectal cancer (48 in T1 stage, 127 in T2, 259 in T3, and 66 in T4) were randomly divided into a training group (n = 400) and a validation group (n = 100). Regions of interest (ROIs) in rectal cancer lesions were pixel-wise annotated by experienced radiologists. A deep learning algorithm based on regional attention was used to train a binary classification model (early stage-T1 and T2, advanced stage-T3 and T4) and a multi-classification model (T1, T2, T3 and T4 stages), which were compared against radiomics approaches. Features were extracted from manually segmented ROIs using pyradiomics, radiomics-based binary and multi-classification models using ten different algorithms. In addition, baseline clinical data-based binary and multi-classification models were also constructed. The performance of both binary and multi-classification models were evaluated by plotting receiver operating characteristic (ROC) curves. The area under the curve (AUC) and accuracy were calculated for the binary model, and the micro-average AUC, macro-average AUC, and accuracy were calculated for the multi-classification model. Results: The ROI-based binary classification model for T stage (ROITransStage; AUC = 0.878, accuracy = 0.850) outperformed the best among ten radiomics-based binary models (AdaBoost; AUC = 0.802, accuracy = 0.76), as well as the best-performing baseline clinical data binary model (AdaBoost; AUC = 0.836, accuracy = 0.76). In addition, ROITransStage (micro-average AUC = 0.873, macro-average AUC = 0.862, accuracy = 0.81) also demonstrated superior diagnostic performance for the T1, T2, T3 and T4 stages compared to the best-performing radiomics-based (SVM; micro-average AUC = 0.845, macro-average AUC = 0.777, accuracy = 0.6) and baseline clinical data-based (SVM; micro-average AUC = 0.841, macro-average AUC = 0.76, accuracy = 0.61) multi-classification models. Conclusions: The CT deep learning binary and multi-classification models based on regional attention exhibited superior predictive performance for rectal cancer staging compared to both radiomics and clinical data-based models.