Differentiating diffuse cystic lung disease and emphysema cases from normal using Artificial Intelligence

Abstract

Introduction: Diffuse Cystic Lung Disease (DCLD) share a common phenotype of multiple thin-walled pulmonary cysts. Their relative scarcity and visual similarity to more prevalent diseases like emphysema, gives rise to frequent misdiagnosis, leading to dramatically worse clinical outcomes and a higher burden on the healthcare system. Aims and Objectives: To stratify DCLD and Emphysema from those with no disease (Normal) using Artificial Intelligence (AI) techniques applied to CT images. Methods: Deep learning models to stratify DCLD and Emphysema patients from normal patients were employed in two studies. The first study isolated the lungs and trained three CNN classifiers (DenseNet201, ResNet50, and Xception) for comparison. The second study trained a ResNet50 model with a Convolutional Block Attention Module to assess potential benefits of using attention components. Explainability was explored using GradCam. Results: The first study used 118 DCLD and control cases, and the second study used 20 DCLD, emphysema and control cases. These datasets yielded 17,460 and 5,312 CT slices respectively. Train and test datasets were created with a 90/10 and 75/25 split respectively. Validation sets were derived from training data during training. Care was taken to ensure that there was no patient overlap between training, validation and testing datasets. AUCs above 0.97 and 0.921, and Average PR above 0.95 and 0.93 were observed across the two studies respectively. Conclusions: The high specificity and AUC scores achieved by the models supports its use case as a decision support tool for radiologists in the analysis of Emphysema and DCLD cases.