HKUST Develops Biomedical AI Model With Diagnostic Performance and Explainability

Researchers at the Hong Kong University of Science and Technology developed ConceptCLIP, a biomedical foundation model that combines strong diagnostic performance with explanations at the concept level that clinicians can understand.

A team led by Professor Chen Hao at the Hong Kong University of Science and Technology published the study in the international journal Nature Biomedical Engineering on the 17th.

Artificial intelligence (AI) used in medical imaging must combine expert-level diagnostic accuracy with explainability that supports clinical decision-making. However, existing multimodal biomedical foundation models have prioritized performance and operated like “black boxes,” failing to provide the detailed, clinically meaningful concept-based explanations needed in clinical practice.

To address this problem, the researchers built MedConcept-23M, a large-scale dataset comprising 23 million biomedical image-text-concept triplets. They pretrained ConceptCLIP on this dataset by jointly aligning images with text and regions with concepts, enabling accurate and explainable medical image analysis.

ConceptCLIP demonstrated strong diagnostic performance while providing human-understandable explanations in a large-scale benchmark spanning 10 imaging modalities and 78 datasets.

In a clinician user study covering three imaging modalities, the concept-based explanations provided by ConceptCLIP helped clinicians validate the model’s predictions and identify potential errors.

The researchers said ConceptCLIP could be a key milestone toward the broad clinical adoption of AI.