Deformation-aided lung segmentation: Introducing the DALS network for efficient learning

• Multitasking architecture leveraging image registration to improve lung segmentation • Efficient training with reduced need for annotations • Indirectly supervised training of segmentation • Unsupervised training of registration and of motion mask extraction Non-invasive regional lung function assessment holds promise for personalized treatment planning in critical care settings. However, to assess lung parenchymal aeration from 3D computed tomography scans, the lungs must be isolated from other organs. This task is difficult in acute respiratory distress syndrome, where extensive and heterogeneous consolidations suppress contrasts within the lungs and between the lung parenchyma and surrounding tissues. Despite the promising performance of deep learning models, the high amount of annotated data required to train such models remains an issue. Here, we propose a multitasking deep neural network, in which lung segmentation in pairs of 3D scans acquired at two different volumes benefits from the estimated deformation field between the two scans. Our learning framework limits the need for annotations: deformation learning is unsupervised, while segmentation learning requires annotations – reference lung masks – only for one scan per pair. We trained our method on moderate to severe ARDS cases without specific preprocessing for noise or artifacts. Evaluated on sixteen challenging cases and compared to three state-of-the-art models, it demonstrated superior performance, while using half as many annotations for training. Although the strongest competitor achieved, on average, the same overlap with the reference lung masks (mean Dice score 0.931), our model generated significantly fewer outliers (mean Hausdorff distance from reference mask surface 26.4mm vs. 50.5 mm).