Confidence Estimation in Analyzing Intravascular Optical Coherence Tomography Images with Deep Neural Networks

Atherosclerosis is a leading, yet preventable cause of death worldwide. To diagnose and assess plaque deposits within the arterial walls, experts employ intravascular optical coherence tomography (IVOCT) for the detection and characterization of lesions. Clinical routines acquire a high number of images, necessitating automatic plaque detection for fast and accurate decision support. Deep learning-based plaque detection methods demonstrated remarkable success, but come with frequent prediction inaccuracies. Therefore, a thorough understanding and assessment of predictive reliability is required for future clinical decision support systems. We address this issue by employing a novel uncertainty estimation method, extending deep learning (DL) models that directly learn plaque classification from an IVOCT dataset and form the basis of state-of-the-art architectures, through a confidence estimation branch (CEB). Our models allow for optimized treatment execution by providing physicians with confidence scores, along with disease predictions from DL models. Experiments demonstrate that the estimated confidence correlates to prediction accuracy with the proposed method. We show that generating uncertainty scores with a CEB is feasible, paving the way for more qualitative and effective decision-making in future medical predictive analysis.