Geometric Heuristics Enhance POCUS AI for Pneumothorax

Abstract

Point-of-care ultrasound (POCUS) represents a significant advancement in emergency and critical care medicine, offering real-time imaging capabilities that are essential for rapid diagnostic and therapeutic decisions. Despite its advantages, the effective interpretation of POCUS images demands a high degree of expertise, making the process susceptible to variability and potential diagnostic inaccuracies. POCUS can also be challenging for artificial intelligence (AI) to interpret, and existing commercial POCUS AI systems required large amounts of labeled training data. We trained POCUS AI models using limited training data, for which we improved diagnostic accuracy by directly integrating geometric heuristics into the architecture and training of our POCUS AI models. These heuristics were derived by humans from the expert knowledge of clinicians. Our clinical focus was diagnosing pneumothorax, where rapid and accurate detection is crucial. Our AI-enhancing methodology centered on collecting and incorporating geometric heuristics, i.e. intuitive rules and patterns used by clinicians in image interpretation. These heuristics include the observation of pleural line sliding, its relative movement against the intercostal muscle, and the specific positioning of the pleural line, among others. We represented these heuristics using semantic-segmentation label images and optical flow images, and we also cropped the images based on the semantic segmentation. We developed two distinct methods for embedding these additional heuristic images into the AI models: one through adding new channels to the input data and another by integrating them as distinct inputs that the AI model later fuses into a common embedding space with the original grayscale image data. The strategies of cropping regions of interest (ROI) and utilizing segmentation maps significantly outperformed the baseline models, underscoring the importance of directing the AI’s focus to crucial image regions. Surprisingly, optical flow maps did not enhance model performance, highlighting the nuanced nature of computing and/or integrating motion-related heuristics for ultrasound. Overall, the multi-channel input approach proved slightly more effective than the fused embedding space, though both methods showed promise in improving diagnostic accuracy. This study opens up avenues for exploring additional heuristic combinations and refining model architectures to further enhance the performance v of AI in medical imaging, especially in applications such as POCUS video for which both clinicians and AI often struggle to discern nondescript anatomy and motion.