Quantification of lung ventilation defects on hyperpolarized MRI: The Multi-Ethnic Study of Atherosclerosis (MESA) COPD study

Abstract

Purpose: To develop an end-to-end deep learning (DL) framework to segment ventilation defects on pulmonary hyperpolarized MRI.

Materials and methods: The Multi-Ethnic Study of Atherosclerosis Chronic Obstructive Pulmonary Disease (COPD) study is a nested longitudinal case-control study in older smokers. Between February 2016 and July 2017, 56 participants (age, mean ± SD, 74 ± 8 years; 34 men) underwent same breath-hold proton (¹H) and helium (³He) MRI, which were annotated for non-ventilated, hypo-ventilated, and normal-ventilated lungs. In this retrospective DL study, 820 ¹H and ³He slices from 42/56 (75%) participants were randomly selected for training, with the remaining 14/56 (25%) for test. Full lung masks were segmented using a traditional U-Net on ¹H MRI and were imported into a cascaded U-Net, which were used to segment ventilation defects on ³He MRI. Models were trained with conventional data augmentation (DA) and generative adversarial networks (GAN)-DA.

Results: Conventional-DA improved ¹H and ³He MRI segmentation over the non-DA model (P = 0.007 to 0.03) but GAN-DA did not yield further improvement. The cascaded U-Net improved non-ventilated lung segmentation (P < 0.005). Dice similarity coefficients (DSC) between manually and DL-segmented full lung, non-ventilated, hypo-ventilated, and normal-ventilated regions were 0.965 ± 0.010, 0.840 ± 0.057, 0.715 ± 0.175, and 0.883 ± 0.060, respectively. We observed no statistically significant difference in DCSs between participants with and without COPD (P = 0.41, 0.06, and 0.18 for non-ventilated, hypo-ventilated, and normal-ventilated regions, respectively).

Conclusion: The proposed cascaded U-Net framework generated fully-automated segmentation of ventilation defects on ³He MRI among older smokers with and without COPD that is consistent with our reference method.

Keywords: COPD; Deep learning; Hyperpolarized gas; MRI; Ventilation defects.

Fig. 1.

Illustration of ground truth segmentation. Full lung masks were segmented on ¹H MRI and subsequently applied to ³He MRI. To determine appropriate thresholds for full lung masks and ventilation segmentations, ROIs were manually drawn inside the lungs on ¹H MRI, and inside the central airways and the heart on ³He MRI. SD = standard deviation; SI = signal intensity; ROI = region of interest. Colour coding: Orange = central airway ROI; blue = heart ROI; red = non-ventilated lung regions; yellow = hypo-ventilated lung regions; green = normal-ventilated lung regions. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Fig. 2.

Conv2d = 2D convolution, BN = batch normalization, LeakyReLU = leaky rectified linear unit, ConvTrans2d = 2D transpose convolution. a): U-Net architecture, in which we substituted the max pooling layer with a 2D convolution layer (Conv2d) with a 4*4 kernel. b): The proposed end-to-end deep learning framework. Full lung mask was segmented from ¹H MRI through a traditional U-Net. For ³He segmentation, we applied a two-layer cascaded U-Net which first segmented ³He MRI into ventilated (blue) and non-ventilated (red) lung regions. The ventilated regions were then segmented into normal-(green) and hypo-(yellow) ventilated lung regions. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Fig. 3.

PG = progressively-growing. Data augmentation (DA). a): Conventional-DA includes random transformations: scaling (scaled by 80–120% along each axis separately), translation (translated by −10 to 10% pixels per axis), rotation (−45° to 45 °), shearing (−15 ° to 15 °), horizontal flip (probability P = 0.5), vertical flip (probability P = 0.5); b): GAN-based DA used two separate generative adversarial networks (GANs). The first unconditional GAN generated random synthetic full lung masks, and the second conditional GAN translated the random masks into corresponding synthetic ¹H MR image.

Fig. 4.

Comparison of ventilation segmentation results between traditional U-Net and cascaded U-Net. Red = non-ventilated lung region; yellow = hypo-ventilated lung region; green = normal-ventilated lung regions. The circled regions highlight higher agreement with cascaded U-Net segmentation than with traditional U-Net segmentation when compared to the ground truth. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)