Title: Enhancing Medical Image Diagnosis with Label-Efficient Interpretability through a Semi-Supervised Hypergraph Concept Bottleneck Model

Abstract:

While deep learning has transformed medical image analysis by achieving remarkable diagnostic precision across various fields, its widespread clinical integration is often obstructed by a lack of transparency. In critical healthcare scenarios, where trust relies heavily on understanding decision-making processes, this opacity is a significant barrier. A prime example is the diagnosis of Placenta Accreta Spectrum (PAS); here, ultrasound images contain faint indicators that make accurate scoring difficult, causing uninterpretable "black-box" models to be deemed unreliable.

Concept Bottleneck Models (CBMs) present a potential solution by incorporating clinically relevant intermediate concepts into the diagnostic workflow. This structure allows medical professionals to review and adjust the model’s conclusions. However, traditional CBMs struggle to account for intricate relationships between concepts and require expensive, expert-level annotations, which restricts their practical application.

To overcome these limitations, this research proposes a new semi-supervised CBM framework tailored for medical imaging. The method utilizes dual-level hypergraph learning to capture high-order dependencies among concepts and to create domain-adaptive pseudo-labels. By combining a concept-level hypergraph for improved reasoning with an image-level hypergraph for stable pseudo-label creation, the proposed method delivers both high performance and clear interpretability. The framework’s efficacy is demonstrated through experiments on a newly annotated PAS ultrasound dataset and a public breast ultrasound dataset. Additionally, its broad applicability is confirmed using the SkinCon dermoscopic image dataset. The source code for this project can be accessed at https://github.com/scott-yjyang/HyperCBM.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC