Evaluating UNet Performance in Segmenting Cysts of Diffuse Cystic Lung Disease CT Scans

Diffuse Cystic Lung Diseases (DCLDs) are characterised by the presence of cysts in the parenchyma of the lungs. Lymphangioleiomyomatosis (LAM) is a DCLD characterised generally by regular-shaped cysts. Accurately isolating these cysts in chest CTs for further quantification can facilitate assessment of disease burden and treatment response. Our research aims to develop machine learning and artificial intelligence techniques (ML/AI) to reliably isolate cysts in DCLD CT scans to facilitate assessment of their distribution, shape, size and cyst volume percent. A team of radiologists manually segmented cysts present in 22 high-resolution chest CTs from clinically confirmed LAM cases with varying scan quality and disease burden. Samples were split into patient independent datasets with a 70/30 train/test split. A deep-learning segmentation model (UNet) was trained on the 1481 axial CT images of the training subset to segment cysts. The model's performance was assessed by comparing manual vs. predicted cyst segmentation on the axial CT Images of the testing subset (558 Images). The assessments yielded an average DICE Score of 51.8%, precision of 68.2% and sensitivity of 46.2% for isolating cysts. Results are impacted by cyst variation. There is potential for ML/AI to be used to isolate cysts in Chest CT scans of DCLD cases. This provides a platform to facilitate quantification of cyst metrics, disease burden and treatment responses. <fig><object-id>erj;64/suppl_68/PA5209/F1</object-id><object-id>F1</object-id><object-id>F1</object-id><graphic></graphic></fig>