MIT and collaborators developed a technique to make computer vision models, including those used in medical imaging, provide clearer, concept-based explanations for their predictions.
Key Details
- 1The new method extracts and uses concepts learned by the model during training for explanations, instead of relying solely on human-defined concepts.
- 2A sparse autoencoder and multimodal large language model are used to extract and describe these concepts in plain language.
- 3The approach limits the model to using five concepts per prediction for clarity and relevance.
- 4Compared to state-of-the-art concept bottleneck models (CBMs), this technique achieved higher accuracy and more precise explanations in tasks including medical image diagnosis.
- 5The research will be presented at the International Conference on Learning Representations, with future plans to scale the approach and address information leakage.
Why It Matters
Improved explainability is critical for the adoption and trust of AI models in medical diagnostics, where human-understandable reasoning can support clinicians' decision-making. Techniques that balance accuracy with transparency help meet regulatory and ethical demands in radiology AI.
Source
EurekAlert
Related News
•EurekAlert
NYU AI Model Improves 5-Year Breast Cancer Risk Prediction via 3D Mammograms
NYU researchers developed an AI model using longitudinal 3D mammograms that outperforms single-scan and 2D-based tools in predicting 5-year breast cancer risk.
•EurekAlert
AI Framework Accelerates Aortic Aneurysm Risk Prediction from Imaging
Researchers developed BioPINN-LM, combining physics-informed neural networks and multimodal large language models to deliver fast, interpretable risk assessments for ascending thoracic aortic aneurysms.
•EurekAlert
Study Finds Patient Voices Missing in Generative AI Design for Oncology
A Flinders University-led review found patients and carers are rarely involved in shaping generative AI tools used in oncology.