Hemi-diaphragm detection of chest X-ray images based on convolutional neural network and graphics.

This article focuses on developing an effective hemi-diaphragm detection method for postero-anterior (P-A) chest X-ray (CXR) images to assess diaphragm function in critically ill and emergency patients. The method combines convolutional neural network (CNN)-based lung field segmentation with a novel graphics-based localization of the cardiophrenic angle, leveraging the two-dimensional morphology of the left and right lungs. Using multi-center datasets comprising 789 static and dynamic P-A CXR images, five CNN models were trained and evaluated, with AttU-Net combined with a connected domain algorithm achieving the best segmentation accuracy and hemi-diaphragm localization performance. The mean Euclidean distance errors for key hemi-diaphragm points ranged from approximately 4.43 to 9.05 pixels across models and image types, indicating the method's potential utility for precision healthcare in vulnerable populations. Limitations include reliance on basic CNN architectures and a limited number of dynamic images, with future work encouraged to expand datasets and explore other imaging views.