Li Z, Li R, Kiser KJ, Giancardo L, Zheng WJ. Segmenting thoracic cavities with neoplastic lesions: a head-to-head benchmark with fully convolutional neural networks.

BCB '21: Proceedings of the 12th ACM Conference on Bioinformatics, Computational Biology, and Health Informatics. 2021; Article No. 33:1-8. https://doi.org/10.1145/3459930.3469564

Summary: This study compares four fully convolutional neural network (FCNN) architectures for segmenting thoracic cavities in CT scans, focusing on patients with neoplastic disease. Using a dataset of 402 cancer patients, the models were evaluated based on Dice coefficient, average symmetric surface distance (ASSD), and 95% Hausdorff distance (HD95). The two-stage 3D U-Net slightly outperformed others, with Dice coefficients of 0.947 for the left lung and 0.952 for the right. The 3D U-Net achieved top scores for HD95 (right lung) and ASSD (both lungs). The study highlights the effectiveness of FCNNs in segmenting lungs with cancerous lesions, enabling their use for downstream clinical applications like treatment planning and prognosis.