A radiographic, deep transfer learning framework, adapted to estimate lung opacities from chest x-rays (opens in new tab)
Objective: To develop and validate a deep learning framework for estimating chest X-ray (CXR) lung opacity severity, which could assist radiologists in standardizing opacity assessment. Methods: We developed a transfer learning framework using 38,079 training CXR images and validated against expert radiologist annotations using 286 out-of-sample images. Three neural network architectures (ResNet-50, VGG-16, and ChexNet) were tested with different segmentation and data balancing strategies. Results: ResNet-50 with undersampling and no region-of-interest segmentation provided optimal performance. The model's opacity score predictions showed superior agreement with radiologist scores compared to inter-radiologist agreement. The framework provides automated opacity quantification while maintaining high concordance with expert radiologist assessments.