Exploring Strategies for Identifying and Resolving Challenges in Deep Learning-Based Biventricular Segmentation

The translation of machine learning research findings into clinical practice presents numerous challenges. This paper focuses on identifying critical issues related to this translation process, specifically in the domain of medical image segmentation. We propose strategies to systematically tackle these challenges, with particular attention to scenarios where the model's segmentation results are inaccurate, referred to as 'corner cases.' A widely used metric for analyzing the performance of various segmentation algorithms is the average Dice score of all patients. However, we have seen a significant limitation with this aggregated reporting approach, as it often fails to highlight corner cases where an algorithm or model exhibits errors or very low metrics. As a result, models that appear to have better overall performance may produce anatomically erroneous results in several challenging cases without detection. To illustrate this issue, we demonstrate how corner cases can go unnoticed in Automated Cardiac Diagnosis Challenge (ACDC) dataset. To address this limitation, we introduce a framework for identifying and reporting corner cases. Additionally, we propose a new balanced control point scheme aimed at achieving superior performance, even in corner cases. A significant improvement of 2% for the Left Ventricle (LV) segmentation, 1.4% for Right Ventricle (RV) segmentation, and 1.7% for myocardium segmentation has been achieved using our proposed method, in the identified corner case of the ACDC segmentation challenge.