Video Summarization and Fracture Detection in Pediatric Wrist Ultrasound Using Deep Reinforcement Learning

Abstract

Wrist fractures are common injuries among children, significantly impacting daily activities and contributing to prolonged wait times in emergency departments. Portable ultrasound is a safe, radiation-free, and cost-effective diagnostic tool with real-time imaging capabilities to identify these fractures. Ultrasound (US) videos offer dynamic, detailed views for fracture detection but can be lengthy and redundant, making interpretation time-consuming for clinicians. To address this challenge, we developed a novel video summarization and fracture detection method using Deep Reinforcement Learning (DRL). The DRL agent emulates a human expert by analyzing the entire video, selecting diagnostically relevant frames, and leveraging a CNN to classify these frames as either normal or fractured. The reward mechanism, which prioritizes feature similarity and frame dissimilarity, enhances the agent's ability to identify the keyframes. Anisotropic diffusion is applied to US images to enhance bright bony regions, before feature extraction and representativeness calculation. Frame similarity calculations are parallelized to reduce computational complexity, storage demands and simplify the classification task by focusing solely on frames indicative of fractures. On a dataset of 114 patients, the proposed classification network achieved an accuracy of 88.8% with a sensitivity of 92.5% and a specificity of 86.6% using the RL-generated video summaries. Our primary focus is on enhancing sensitivity to ensure that no fractured cases are missed. This performance surpasses the 84.4% accuracy achieved with full video classification. AI-driven ultrasound, with its efficiency, and reduced computational demands, provides a promising solution for early disease detection in resource-constrained environments.Clinical relevance- This AI-driven fast, cost-effective ultrasound tool achieves high accuracy and sensitivity while reducing training time by 50%, making it suitable for real-time clinical use on low-power devices. This can assist lightly-trained healthcare providers in early disease detection, reduce long wait times, and enhance treatment access in remote areas.

Ähnliche Arbeiten