Clinical Evaluation of Accelerated Diffusion-Weighted Imaging of Rectal Cancer Using a Denoising Neural Network

Restaging rectal MRI following total neoadjuvant therapy (TNT) is critical for assessing residual disease and guiding treatment decisions. Diffusion-weighted imaging (DWI) is a key component of rectal MRI protocols, providing functional information that complements anatomical sequences. However, standard DWI acquisitions require multiple repetitions to achieve adequate signal-to-noise ratio (SNR), resulting in prolonged scan times that can be uncomfortable for patients and limit clinical throughput. Accelerating DWI by reducing the number of repetitions inherently decreases SNR, potentially compromising image quality and diagnostic accuracy. Deep learning denoising techniques offer a promising solution by removing noise from accelerated acquisitions, thereby restoring image quality. This study evaluates the effectiveness of a convolutional neural network (CNN) trained to denoise accelerated DWI acquisitions in patients with locally advanced rectal cancer undergoing restaging MRI after TNT. The primary objectives were to assess image quality, lesion visibility, and diagnostic accuracy of denoised accelerated DWI compared to standard DWI, and to evaluate inter-reader agreement.

A total of 46 patients (median age, 60 years [IQR: 47–72]; 37 men and 9 women) were included. The 8- and 16-fold accelerated images, after denoising, maintained or exhibited enhanced lesion visibility and image quality compared to original images acquired with 16 repetitions. Denoised images maintained diagnostic accuracy, with conditional specificities of up to 96%. Inter-reader agreement was moderate to high, indicating reliable image and diagnostic assessment. The overall test yield for denoised DWI reconstructions ranged from 76–98%, demonstrating a reduction in equivocal interpretations. These results suggest that denoising enables substantial acceleration without compromising clinical performance.