The role of deep learning in structural and functional lung imaging

Background: Structural and functional lung imaging are critical components of pulmonary patient care. Image analysis methods, such as image segmentation, applied to structural and functional lung images, have significant benefits for patients with lung pathologies, including the computation of clinical biomarkers. Traditionally, machine learning (ML) approaches, such as clustering, and computational modelling techniques, such as CT-ventilation imaging, have been used for segmentation and synthesis, respectively. Deep learning (DL) has shown promise in medical image analysis tasks, often outperforming alternative methods. Purpose: To address the hypothesis that DL can outperform conventional ML and classical image analysis methods for the segmentation and synthesis of structural and functional lung imaging via: i. development and comparison of 3D convolutional neural networks (CNNs) for the segmentation of ventilated lung using hyperpolarised (HP) gas MRI. ii. development of a generalisable, multi-centre CNN for segmentation of the lung cavity using 1H-MRI. iii. the proposal of a framework for estimating the lung cavity in the spatial domain of HP gas MRI. iv. development of a workflow to synthesise HP gas MRI from multi-inflation, non-contrast CT. v. the proposal of a framework for the synthesis of fully-volumetric HP gas MRI ventilation from a large, diverse dataset of non-contrast, multi-inflation 1H-MRI scans. Methods: i. A 3D CNN-based method for the segmentation of ventilated lung using HP gas MRI was developed and CNN parameters, such as architecture, loss function and pre-processing were optimised. ii. A 3D CNN trained on a multi-acquisition dataset and validated on data from external centres was compared with a 2D alternative for the segmentation of the lung cavity using 1H-MRI. iii. A dual-channel, multi-modal segmentation framework was compared to single-channel approaches for estimation of the lung cavity in the domain of HP gas MRI. iv. A hybrid data-driven and model-based approach for the synthesis of HP gas MRI ventilation from CT was compared to approaches utilising DL or computational modelling alone. v. A physics-constrained, multi-channel framework for the synthesis of fully-volumetric ventilation surrogates from 1H-MRI was validated using five-fold cross-validation and an external test data set. Results: i. The 3D CNN, developed via parameterisation experiments, accurately segmented ventilation scans and outperformed conventional ML methods. ii. The 3D CNN produced more accurate segmentations than its 2D analogues for the segmentation of the lung cavity, exhibiting minimal variation in performance between centres, vendors and acquisitions. iii. Dual-channel, multi-modal approaches generate significant improvements compared to methods which use a single imaging modality for the estimation of the lung cavity. iv. The hybrid approach produced synthetic ventilation scans which correlate with HP gas MRI. v. The physics-constrained, 3D multi-channel synthesis framework outperformed approaches which did not integrate computational modelling, demonstrating generalisability to external data. Conclusion: DL approaches demonstrate the ability to segment and synthesise lung MRI across a range of modalities and pulmonary pathologies. These methods outperform computational modelling and classical ML approaches, reducing the time required to adequately edit segmentations and improving the modelling of synthetic ventilation, which may facilitate the clinical translation of DL in structural and functional lung imaging.