A comparative study of magnetic resonance imaging image-based deep learning and conventional turbo spin echo sequence in tumor-node staging of rectal cancer.

Background: Accurate tumor-node (TN) staging of rectal cancer is crucial for treatment planning. Conventional turbo spin echo (TSE) magnetic resonance imaging (MRI) is time-consuming and limited by image quality. This study aimed to compare the diagnostic performance, staging accuracy, and acquisition time of deep learning-reconstructed TSE (DL TSE) with conventional TSE (conv TSE) in rectal cancer MRI. Methods: From October 2023 to October 2024, 60 patients with rectal tumors were prospectively enrolled following diagnosis at the Cancer Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College. All participants underwent MRI scanning and TN staging by DL TSE and conv TSE, and a comparison was made using the McNemar's test and Kappa test. Results: In TN staging diagnosis, DL TSE sequences performed significantly better than conv TSE sequences. By applying DL reconstruction to T1-weighted imaging (T1WI) and T2-weighted imaging (T2WI) TSE sequences, T staging accuracy was improved from 0.778 to 0.907, and N staging accuracy was improved from 0.703 to 0.926. In the T-stage assessment, McNemar's test showed for conv TSE (χ Conclusions: The accuracy, sensitivity, and specificity of DL TSE sequences for rectal cancer TN staging were higher than those of conv TSE. The DL TSE sequence (including T1WI and T2WI) results for TN staging of rectal cancer were close to the pathological diagnosis, indicating a high clinical application for TN staging of rectal cancer. DL TSE can improve diagnostic accuracy and reduce scanning time by 36.6%.