Implementable Deep Learning for Multi‐sequence Proton MRI Lung Segmentation: A Multi‐center, Multi‐vendor, and Multi‐disease Study

Abstract

BackgroundRecently, deep learning via convolutional neural networks (CNNs) has largely superseded conventional methods for proton (1H)‐MRI lung segmentation. However, previous deep learning studies have utilized single‐center data and limited acquisition parameters.PurposeDevelop a generalizable CNN for lung segmentation in 1H‐MRI, robust to pathology, acquisition protocol, vendor, and center.Study typeRetrospective.PopulationA total of 809 1H‐MRI scans from 258 participants with various pulmonary pathologies (median age (range): 57 (6–85); 42% females) and 31 healthy participants (median age (range): 34 (23–76); 34% females) that were split into training (593 scans (74%); 157 participants (55%)), testing (50 scans (6%); 50 participants (17%)) and external validation (164 scans (20%); 82 participants (28%)) sets.Field Strength/Sequence1.5‐T and 3‐T/3D spoiled‐gradient recalled and ultrashort echo‐time 1H‐MRI.Assessment2D and 3D CNNs, trained on single‐center, multi‐sequence data, and the conventional spatial fuzzy c‐means (SFCM) method were compared to manually delineated expert segmentations. Each method was validated on external data originating from several centers. Dice similarity coefficient (DSC), average boundary Hausdorff distance (Average HD), and relative error (XOR) metrics to assess segmentation performance.Statistical TestsKruskal–Wallis tests assessed significances of differences between acquisitions in the testing set. Friedman tests with post hoc multiple comparisons assessed differences between the 2D CNN, 3D CNN, and SFCM. Bland–Altman analyses assessed agreement with manually derived lung volumes. A P value of <0.05 was considered statistically significant.ResultsThe 3D CNN significantly outperformed its 2D analog and SFCM, yielding a median (range) DSC of 0.961 (0.880–0.987), Average HD of 1.63 mm (0.65–5.45) and XOR of 0.079 (0.025–0.240) on the testing set and a DSC of 0.973 (0.866–0.987), Average HD of 1.11 mm (0.47–8.13) and XOR of 0.054 (0.026–0.255) on external validation data.Data ConclusionThe 3D CNN generated accurate 1H‐MRI lung segmentations on a heterogenous dataset, demonstrating robustness to disease pathology, sequence, vendor, and center.Evidence Level4.Technical EfficacyStage 1.

A Dual‐Channel Deep Learning Approach for Lung Cavity Estimation From Hyperpolarized Gas and Proton MRI

“…29 Previous studies have described DL-based approaches to segment the lung parenchyma on 1 H-MR images; however, these approaches have conducted the segmentation using single-channel networks. [16][17][18] The inclusion of functional features present in the hyperpolarized gas MRI scans may provide the network context with which to adapt the structural LCE to account for inherent registration errors between the 1 H-MRI and 129 Xe-MRI acquisitions. Previous work by Tustison et al utilized separate networks for segmenting 1 H-MRI and hyperpolarized gas MRI 16 ; however, due to several factors, including inherent registration errors and differences in inflation levels, a network that generates a structural segmentation purely using 1 H-MRI seems inadequate.…”

“…18 A 3D UNet was employed and achieved a mean DSC of 0.96 for whole-lung segmentation across all resolutions. 18 All these approaches to generate whole-lung segmentations from 1 H-MRI have used single-channel, mono-modal CNN-based methods, where a single image or 3D scan is used as an input to the CNN. [16][17][18] Although these methods have shown promising results, they cannot account for the aforementioned spatial misalignments between structural and functional modalities.…”

“…18 All these approaches to generate whole-lung segmentations from 1 H-MRI have used single-channel, mono-modal CNN-based methods, where a single image or 3D scan is used as an input to the CNN. [16][17][18] Although these methods have shown promising results, they cannot account for the aforementioned spatial misalignments between structural and functional modalities. Multichannel approaches using multimodal images have shown promise in DL image analysis applications, where there are important features across multiple imaging modalities.…”

“…Astley et al have demonstrated accurate 1 H‐MRI segmentation on a large dataset, containing multi‐resolution scans of patients with various pulmonary pathologies 18 . A 3D UNet was employed and achieved a mean DSC of 0.96 for whole‐lung segmentation across all resolutions 18 . All these approaches to generate whole‐lung segmentations from 1 H‐MRI have used single‐channel, mono‐modal CNN‐based methods, where a single image or 3D scan is used as an input to the CNN 16–18 .…”

“…Zha et al used a two‐dimensional (2D) UNet to successfully segment ultra‐short echo time (UTE) 1 H‐MRI scans; however, this work used a relatively limited dataset, containing only 45 participants 17 . Astley et al have demonstrated accurate 1 H‐MRI segmentation on a large dataset, containing multi‐resolution scans of patients with various pulmonary pathologies 18 . A 3D UNet was employed and achieved a mean DSC of 0.96 for whole‐lung segmentation across all resolutions 18 .…”