Using Radiomic Features to Detect Anatomical Errors and Assess Deep Learning-Based Left Ventricle Segmentation in Cardiac MRI.
Authors
Affiliations (2)
Affiliations (2)
- University of São Paulo, São Paulo, São Paulo, Brazil. [email protected].
- University of São Paulo, São Paulo, São Paulo, Brazil.
Abstract
Segmentation of the left ventricle in cardiac magnetic resonance exams is critical for accurate diagnosis and plays a central role in computer-aided diagnosis systems. Among the various automatic methods proposed, deep learning-based approaches have demonstrated promising results, often achieving segmentation quality comparable to that of expert annotations. However, a significant limitation of these methods is their susceptibility to anatomical inconsistencies that may occur arbitrarily and remain undetected by conventional evaluation metrics, thereby compromising clinical interpretation. To address this issue, we explore a novel use of Radiomics to assess the anatomical quality of deep learning segmentations. Specifically, we extract radiomic features from deep learning-based left ventricle segmentations and use them to train machine learning classifiers aimed at identifying segmentations with anatomical errors. We perform extensive cross-validation experiments on five public and one private datasets and analyze classification performance considering different severity levels of anatomical errors. Experimental results indicate that the proposed radiomic-based classifiers achieved very high classification performance (Accuracy, Recall, and Specificity > 95% in most cases). The classifiers were able to detect errors (Recall > 80%) even in less severe cases, where standard metrics such as the Dice similarity coefficient indicate good segmentation quality (Dice > 0.8). Inter-dataset analysis indicates high generalization ability (F1-score > 0.8) and applicability on data from different clinical settings. The results demonstrate that radiomic-based classifiers can be reliably used for automatically detecting anatomical errors in segmentations, reducing the need for ground truth annotations and offering an alternative for quality control assessment of segmentation approaches.