To explore the impact of deep learning reconstruction (DLR) technology on image quality and the diagnostic performance of preoperative rectal cancer.

A retrospective analysis was conducted on patients with rectal cancer confirmed by postoperative pathology in our hospital from July 2017 to March 2022, focusing on their clinical, pathological, and pre-treatment MRI data. The DLR technology was used to reconstruct and optimize the T₂WI, DWI, and CE-T₁WI sequences in MRI. Five readers with different diagnostic experiences separately assessed the image quality of the original-MRI and DLR-MRI image sets. Objective quantitative evaluation of the images was conducted by calculating the signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR) of the lesions, while subjective qualitative analysis was performed on overall image quality and lesion display performance. Finally, subjective diagnosis of preoperative TN staging and comparison of diagnostic performance of deep learning models were performed for the two groups of images.

A total of 178 patients were included. The DLR-MRI images showed significantly higher values in both objective quantitative metrics (SNR and CNR) and subjective evaluations (overall image quality and lesion display performance) compared to the original images (P<0.0001). In terms of the diagnostic efficacy of MRI images, DLR-MRI increased the AUC values of preoperative T staging for rectal cancer in readers 3 and 5, reaching up to 0.921 (reader 3), but the difference was not statistically significant (P>0.05). Among them, the diagnostic efficacy of reader 4 was significantly improved (Z=2.971, P=0.003). Moreover, DLR-MRI significantly enhanced the diagnostic efficacy of the deep learning classification model, with the AUC value increasing from 0.797 to 0.937 (Z=2.505, P=0.043). However, neither the subjective assessment nor the DL classification model achieved statistically significant improvements in preoperative N-staging for rectal cancer. Further analysis using DCA showed that in preoperative T-staging of rectal cancer, the DLR-DL model demonstrated higher clinical net benefit, while the clinical benefits of the two models were similar in preoperative staging.

DLR technology can optimize MRI image quality to enhance lesion display capabilities. Furthermore, the deep learning classification model based on DLR can improve the preoperative diagnostic performance of rectal cancer to a certain extent.