Large-scale investigation of deep learning approaches for ventilated lung segmentation using multi-nuclear hyperpolarized gas MRI

Abstract

Respiratory diseases are leading causes of mortality and morbidity worldwide. Pulmonary imaging is an essential component of the diagnosis, treatment planning, monitoring, and treatment assessment of respiratory diseases. Insights into numerous pulmonary pathologies can be gleaned from functional lung MRI techniques. These include hyperpolarized gas ventilation MRI, which enables visualization and quantification of regional lung ventilation with high spatial resolution. Segmentation of the ventilated lung is required to calculate clinically relevant biomarkers. Recent research in deep learning (DL) has shown promising results for numerous segmentation problems. Here, we evaluate several 3D convolutional neural networks to segment ventilated lung regions on hyperpolarized gas MRI scans. The dataset consists of 759 helium-3 (3He) or xenon-129 (129Xe) volumetric scans and corresponding expert segmentations from 341 healthy subjects and patients with a wide range of pathologies. We evaluated segmentation performance for several DL experimental methods via overlap, distance and error metrics and compared them to conventional segmentation methods, namely, spatial fuzzy c-means (SFCM) and K-means clustering. We observed that training on combined 3He and 129Xe MRI scans using a 3D nn-UNet outperformed other DL methods, achieving a mean ± SD Dice coefficient of 0.963 ± 0.018, average boundary Hausdorff distance of 1.505 ± 0.969 mm, Hausdorff 95th percentile of 5.754 ± 6.621 mm and relative error of 0.075 ± 0.039. Moreover, limited differences in performance were observed between 129Xe and 3He scans in the testing set. Combined training on 129Xe and 3He yielded statistically significant improvements over the conventional methods (p < 0.0001). In addition, we observed very strong correlation and agreement between DL and expert segmentations, with Pearson correlation of 0.99 (p < 0.0001) and Bland–Altman bias of − 0.8%. The DL approach evaluated provides accurate, robust and rapid segmentations of ventilated lung regions and successfully excludes non-lung regions such as the airways and artefacts. This approach is expected to eliminate the need for, or significantly reduce, subsequent time-consuming manual editing.

A hybrid model‐ and deep learning‐based framework for functional lung image synthesis from multi‐inflation CT and hyperpolarized gas MRI

“…The hybrid DL configuration generated a median DSC of 0.95 for ventilated regions and 0.48 for defect regions, significantly outperforming the DSC achieved by the CT HU method. Both VLP values and regional overlap values require the segmentation of synthetic ventilation scans and are, therefore, susceptible to biases in the segmentation algorithm used; the automatic segmentation method used here was trained to segment hyperpolarized gas MRI and not synthetic ventilation scans 52 . There is limited consensus on the appropriate segmentation schema required for the delineation of ventilated and defect regions, resulting in an inability to produce accurate comparisons between research studies.…”

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

“…The augmentation method did not increase the total size of the dataset but instead used random rotation and scaling factors to modify scans before entering the network. Rotation angles of −10° to 10° and scaling values of −10% to 10% were applied for each epoch, selected based on previous research investigations 23 . Augmentation techniques were constrained to the above limits to produce physiologically plausible scans.…”

Optimized quantitative mapping of cardiopulmonary oscillations using hyperpolarized 129Xe gas exchange MRI: Digital phantoms and clinical evaluation in CTEPH

“…Thus, the image SNR was calculated as $$$$ SN{R}_{image}=\frac{Mea{n}_{signal}}{Median\left( St{d}_{noise\ per\ window}\right)} $$$$ where the mean signal was that within the phantom/thoracic cavity mask, and the noise was that outside the phantom/thorax. The phantom mask can be trivially calculated using thresholding, whereas thoracic cavity masks for in vivo images were produced using a V‐net‐based convolutional neural network 25,26 …”

“…The phantom mask can be trivially calculated using thresholding, whereas thoracic cavity masks for in vivo images were produced using a V-net-based convolutional neural network. 25,26 Oscillation images were rendered into color maps via linear binning based on the healthy cohort distribution as described in applications for conventional gas exchange MRI. 27,28 However, regions where the SNR of the non-keyhole RBC images fell below the computed noise level (the denominator in Eq.…”