Deep learning-based imaging reconstruction for MRI after neoadjuvant chemoradiotherapy for rectal cancer: effects on image quality and assessment of treatment response

Abstract

Purpose: To investigate the effects of deep learning-based imaging reconstruction (DLR) on the image quality of MRI of rectal cancer after chemoradiotherapy (CRT), and its accuracy in diagnosing pathological complete responses (pCR).

Methods: We included 39 patients (men: women, 21:18; mean age ± standard deviation, 59.1 ± 9.7 years) with mid-to-lower rectal cancer who underwent a long-course of CRT and high-resolution rectal MRIs between January 2020 and April 2021. Axial T2WI was reconstructed using the conventional method (MRIconv) and DLR with two different noise reduction factors (MRIDLR30 and MRIDLR50). The signal-to-noise ratio (SNR) of the tumor was measured. Two experienced radiologists independently made a blind assessment of the complete response on MRI. The sensitivity and specificity for pCR were analyzed using a multivariable logistic regression analysis with generalized estimating equations.

Results: Thirty-four patients did not have a pCR whereas five (12.8%) had pCR. Compared with the SNR of MRIconv (mean ± SD, 7.94 ± 1.92), MRIDLR30 and MRIDLR50 showed higher SNR (9.44 ± 2.31 and 11.83 ± 3.07, respectively) (p < 0.001). Compared to MRIconv, MRIDLR30 and MRIDLR50 showed significantly higher specificity values (p < 0.036) while the sensitivity values were not significantly different (p > 0.301). The sensitivity and specificity for pCR were 48.9% and 80.8% for MRIconv; 48.9% and 88.2% for MRIDLR30; and 38.8% and 86.7% for MRIDLR50, respectively.

Conclusion: DLR produced MR images with higher resolution and SNR. The specificity of MRI for identification of pCR was significantly higher with DLR than with conventional MRI.