Complementation‐reinforced network for integrated reconstruction and segmentation of pulmonary gas MRI with high acceleration

Background Hyperpolarized (HP) gas MRI enables the clear visualization of lung structure and function. Clinically relevant biomarkers, such as ventilated defect percentage (VDP) derived from this modality can quantify lung ventilation function. However, long imaging time leads to image quality degradation and causes discomfort to the patients. Although accelerating MRI by undersampling k‐space data is available, accurate reconstruction and segmentation of lung images are quite challenging at high acceleration factors. Purpose To simultaneously improve the performance of reconstruction and segmentation of pulmonary gas MRI at high acceleration factors by effectively utilizing the complementary information in different tasks. Methods A complementation‐reinforced network is proposed, which takes the undersampled images as input and outputs both the reconstructed images and the segmentation results of lung ventilation defects. The proposed network comprises a reconstruction branch and a segmentation branch. To effectively exploit the complementary information, several strategies are designed in the proposed network. Firstly, both branches adopt the encoder‐decoder architecture, and their encoders are designed to share convolutional weights for facilitating knowledge transfer. Secondly, a designed feature‐selecting block discriminately feeds shared features into decoders of both branches, which can adaptively pick suitable features for each task. Thirdly, the segmentation branch incorporates the lung mask obtained from the reconstructed images to enhance the accuracy of the segmentation results. Lastly, the proposed network is optimized by a tailored loss function that efficiently combines and balances these two tasks, in order to achieve mutual benefits. Results Experimental results on the pulmonary HP 129Xe MRI dataset (including 43 healthy subjects and 42 patients) show that the proposed network outperforms state‐of‐the‐art methods at high acceleration factors (4, 5, and 6). The peak signal‐to‐noise ratio (PSNR), structural similarity (SSIM), and Dice score of the proposed network are enhanced to 30.89, 0.875, and 0.892, respectively. Additionally, the VDP obtained from the proposed network has good correlations with that obtained from fully sampled images (r = 0.984). At the highest acceleration factor of 6, the proposed network promotes PSNR, SSIM, and Dice score by 7.79%, 5.39%, and 9.52%, respectively, in comparison to the single‐task models. Conclusion