Abstract:
The interpretation of Point-of-care ultrasound (POCUS) images poses a challenge due to the scarcity of high-quality labelled data for training AI models in the medical domain. To address this limitation, novel methodologies were developed to train POCUS AI models using limited data, integrating geometric heuristics derived from expert clinicians. Focused on diagnosing pneumothorax, heuristics such as pleural line movement and positioning were collected and embedded into AI models through semantic-segmentation labels and optical flow images. Strategies of cropping ROIs and utilizing segmentation maps significantly outperformed baseline models. While optical flow maps showed no enhancement, this study underscores the potential of leveraging heuristic knowledge to improve AI performance in medical imaging, even with limited labelled data availability.

Committee:
Dr. John Galeotti (advisor)
Dr. Deva Kannan Ramanan
Mr. Nishant Thumbavanam Arun