Accelerating Whole-Body Diffusion-weighted MRI with Deep Learning-based Denoising Image Filters

Our plain-English summary of Accelerating Whole-Body Diffusion-weighted MRI with Deep Learning-based Denoising Image Filters, by K. Zormpas-Petridis, N. Tunariu, A. Curcean, et al (Radiology. Artificial Intelligence, 2021).

The study

• Authors: K. Zormpas-Petridis, N. Tunariu, A. Curcean, C. Messiou, S. Curcean, D.J. Collins, J.C. Hughes, Y. Jamin, D.M. Koh, M.D. Blackledge
• Journal: Radiology. Artificial Intelligence
• Published: 2021;3(5):e200279
• Cited by: 33 articles, as counted by NIH iCite
• Read it: PubMed · Publisher · Free full text

Abstract

Purpose: To use deep learning to improve the image quality of subsampled images (number of acquisitions = 1 [NOA1]) to reduce whole-body diffusion-weighted MRI (WBDWI) acquisition times.

Materials And Methods: Both retrospective and prospective patient groups were used to develop a deep learning-based denoising image filter (DNIF) model. For initial model training and validation, 17 patients with metastatic prostate cancer with acquired WBDWI NOA1 and NOA9 images (acquisition period, 2015-2017) were retrospectively included. An additional 22 prospective patients with advanced prostate cancer, myeloma, and advanced breast cancer were used for model testing (2019), and the radiologic quality of DNIF-processed NOA1 (NOA1-DNIF) images were compared with NOA1 images and clinical NOA16 images by using a three-point Likert scale (good, average, or poor; statistical significance was calculated by using a Wilcoxon signed ranked test). The model was also retrained and tested in 28 patients with malignant pleural mesothelioma (MPM) who underwent lung MRI (2015-2017) to demonstrate feasibility in other body regions.

Results: The model visually improved the quality of NOA1 images in all test patients, with the majority of NOA1-DNIF and NOA16 images being graded as either "average" or "good" across all image-quality criteria. From validation data, the mean apparent diffusion coefficient (ADC) values within NOA1-DNIF images of bone disease deviated from those within NOA9 images by an average of 1.9% (range, 1.1%-2.6%). The model was also successfully applied in the context of MPM; the mean ADCs from NOA1-DNIF images of MPM deviated from those measured by using clinical-standard images (NOA12) by 3.7% (range, 0.2%-10.6%).

Conclusion: Clinical-standard images were generated from subsampled images by using a DNIF.Keywords: Image Postprocessing, MR-Diffusion-weighted Imaging, Neural Networks, Oncology, Whole-Body Imaging, Supervised Learning, MR-Functional Imaging, Metastases, Prostate, Lung Supplemental material is available for this article. Published under a CC BY 4.0 license.

2021 by the Radiological Society of North America, Inc.

In plain English

This proof-of-concept study tested a deep learning image filter meant to shorten diffusion-weighted MRI scans. It did not study asbestos exposure. The filter was developed using scans from 17 men with metastatic prostate cancer from 2015 to 2017. It was then tested on 22 patients with advanced prostate cancer, myeloma or advanced breast cancer, scanned in 2019 with a fast protocol of about 5 minutes and a clinical protocol of 22 to 25 minutes. Two radiologists rated most fast, unfiltered images as poor and most filtered and clinical-standard images as average or good. The differences between filtered and unfiltered images were statistically significant. The study also used lung MRI from 28 patients with malignant pleural mesothelioma: the filter was retrained on 20 of them and checked on the other eight. In those eight validation patients, average diffusion values in disease from filtered images differed from clinical-standard images by 3.7% (range 0.2% to 10.6%). The authors concluded the method could cut the diffusion-weighted part of the scan from about 25 to 5 minutes. They called for larger multicenter studies.

We wrote this plain-English summary from the full article with the help of AI, and a second AI model checked each sentence against the article. It is not the authors' wording; the abstract and the article are their own account.

What the study did and found

• Design: Proof-of-concept technical study developing and testing a deep learning denoising filter for MRI. It used retrospective training and validation data, a prospective test set with a blinded two-radiologist image-quality rating, and a retrospective mesothelioma dataset.
• Who was studied: Three groups. (1) 17 men with metastatic prostate cancer (October 2015 to September 2017), split into 14 for training and 3 for validation. (2) 22 consecutive patients scanned May and June 2019 for testing: 17 with advanced prostate cancer (all men), 3 with myeloma (all men) and 2 with advanced breast cancer (all women). (3) 28 patients (4 women, 24 men; age 52 to 85 years) imaged for malignant pleural mesothelioma (February 2015 to November 2017), split into 20 for training and 8 for validation.
• Exposure: No asbestos or fiber exposure was studied. What was tested was a deep learning denoising image filter (DNIF) applied to fast single-acquisition (NOA1) diffusion-weighted MRI images. Its output was compared with clinical-standard images (NOA9 in training, NOA16 in testing, NOA12 in mesothelioma).
• Main results:
  • In the test patients, the fast diffusion-weighted protocol took about 5 minutes and the clinical diffusion-weighted protocol about 22 to 25 minutes. Patients also had other standard MRI sequences.
  • In the 3 prostate cancer validation patients, filtered images scored better than unfiltered fast images on mean-squared error, structural similarity (0.994 vs 0.992) and peak signal-to-noise ratio (55.7 vs 53.2), each with P less than .001.
  • In the 3 validation patients, mean ADC (apparent diffusion coefficient) values in bone disease from filtered images differed from clinical-standard (NOA9) values by an average relative difference of 1.9% (range 1.1% to 2.6%).
  • In the 22 test patients, both radiologists rated most unfiltered fast images as poor and most filtered and clinical-standard (NOA16) images as average or good. Differences were statistically significant in all comparisons: NOA16 vs NOA1, and filtered vs NOA1.
  • Most filtered images were judged clinically usable, unlike most unfiltered fast images. For example, for signal-to-noise on b = 900 images, 0/22 unfiltered images were usable for both readers. The filtered figures were 17/22 (reader 1) and 21/22 (reader 2), and the clinical-standard figures were 21/22 and 19/22.
  • The lung model was retrained from scratch on 20 of the 28 mesothelioma patients. In the other 8 (validation) patients, it improved mean-squared error, structural similarity and peak signal-to-noise ratio in all eight.
  • In the 8 mesothelioma validation patients, mean ADC in disease differed from clinical-standard (NOA12) images by an average of 2.0% (range 0.4% to 8.4%) for unfiltered fast images. The figures were 3.7% (range 0.2% to 10.6%) for the lung model and 4.0% (range 0.1% to 11.2%) for the original whole-body model.
  • In one mesothelioma patient (patient 3), the filtered-image mean ADC differed from the clinical-standard value by about 11%. The authors reported a similar difference for unfiltered images and judged it not due to the filter.
• The authors' conclusion: The authors concluded that deep learning methods such as their quickDWI approach can improve the quality of whole-body diffusion-weighted MRI images from subsampled data. They said this could potentially reduce the diffusion-weighted acquisition time from about 25 minutes to 5 minutes in their test study.
• Limitations the authors state:
  • The authors described this as a proof-of-concept study needing larger multicenter studies to understand the effect on diagnostic accuracy.
  • Whether the filtered measurements are good enough for monitoring treatment response was not tested. Repeat baseline measurements would be needed.
  • In the mesothelioma images, some bias in ADC values occurred in regions of motion, such as near the diaphragm.
  • Because of breathing motion in the chest, the filter trained on whole-body data tended to oversmooth chest images.

Topics

Medical terms in this study, explained in our Lexicon: Diffusion-Weighted MRI, Myeloma, Statistical Significance, Pleural Mesothelioma.

Related studies

• A histomorphological atlas of resected mesothelioma discovered by self-supervised learning from 3446 whole-slide images — F. Seyedshahi, K. Rakovic, et al. (2025)
• Asbestos Detection with Fluorescence Microscopy Images and Deep Learning — C. Cai, T. Nishimura, et al. (2021)
• Construction of Asbestos Slate Deep-Learning Training-Data Model Based on Drone Images — S.C. Baek, K.H. Lee, et al. (2023)
• Combined diffusion-weighted and gadolinium-enhanced MRI can accurately predict the peritoneal cancer index preoperatively in patients being considered for cytoreductive surgical procedures — R.N. Low, R.M. Barone (2012)
• Diffusion-weighted MRI of malignant pleural mesothelioma: preliminary assessment of apparent diffusion coefficient in histologic subtypes — R.R. Gill, S. Umeoka, et al. (2010)
• Malignant pleural mesothelioma: visual assessment by using pleural pointillism at diffusion-weighted MR imaging — J. Coolen, F. De Keyzer, et al. (2015)

Browse return to medical research.