EchoPCBM: adaptively fitting post-hoc concept bottleneck models for echocardiographic diagnosis and explainability
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
Echocardiography is a cornerstone of cardiovascular imaging, offering real-time, non-invasive assessment of cardiac function and structure. However, interpreting echocardiographic videos remains highly dependent on expert knowledge. While deep neural networks (DNNs) achieve high classification accuracy, their lack of interpretability limits clinical utility. Concept Bottleneck Models (CBMs) address this issue by enforcing predictions through a contextually relevant concept layer, but typically require manual concept annotations which are scarce in echocardiography. To overcome this, we introduce EchoPCBM, a post-hoc Concept Bottleneck Model tailored for echocardiography. EchoPCBM integrates EchoCLIP, a fine-tuned variant of CLIP optimized for echocardiographic images and text, enabling robust concept-image pairing. By dynamically adapting the CBM architecture, EchoPCBM enhances explainability while maintaining strong classification performance. Using private data from 20,250 echocardiographic videos spanning 10 standard imaging views, we validate EchoPCBM on view classification, a fundamental task in echocardiography. We demonstrate that EchoPCBM achieves competitive classification accuracy in comparison to black-box models while providing meaningful concept-level insights, thereby improving model transparency in echocardiographic interpretation.
Fields of Study & Subjects
Explainable Artificial Intelligence (XAI)
Artificial Intelligence in Healthcare and Education
Artificial Intelligence in Healthcare
Publication details
- Journal / source
- SPIE Medical Imaging: Image-Guided Procedures, Robotic Interventions, and Modeling
- Pages
- 13–13
- Publication date
- 1 April 2026
- Language
- EN