An Interpretable Object Detection-Based Model for the Diagnosis of Neonatal Lung Diseases Using Ultrasound Images

Abstract

<p>Lung Ultrasound (LUS) has been progressively used for the diagnosis of lung diseases in neonates due to its safety. Discussions with physicians revealed that a system that can detect specific lung features associated with neonatal lung diseases will be more useful than a simple image classification model. Therefore, a seven-class faster Region Proposal-based Convolutional Neural Network (fRCNN) as well as a RetinaNet were trained on lower posterior lung ultrasound videos to detect seven LUS features. Results show that fRCNN achieved a higher mean Average Precision (mAP) of 86.57% with an Intersection over Union (IoU) of 0.2 compared to RetinaNet with 61.15% mAP. A lung sliding feature detection method was proposed to differentiate between Pneumothorax and Normal scans. Using this method, we were able to classify 5 Pneumothorax (PTX) and 6 Normal video cases with 100% accuracy. We also developed a user-friendly GUI model that performs predictions on LUS video scans and outputs a video with the imposed LUS feature detections.</p>