MedCORE Preprint Proposes Criteria-Grounded Framework for Medical Image Diagnosis
arXiv preprint dated October 6 2026 introduces MedCORE a vision-language model that decomposes diagnosis into clinical criteria localizes evidence and models inter-criteria…
- MedCORE structures clinical reasoning by evaluating specific morphological and textural criteria instead of direct image-to-label mapping.
- The abstract reports quantitative results on three imaging datasets and states improvements over several baseline model types.
- As an abstract-only preprint the full methodology limitations and independent validation remain unavailable.
Addressing the Clinical Reasoning Gap
The preprint authors observe that clinical diagnosis relies on systematic evaluation of morphological and textural criteria. Many deep learning models instead map image features directly to disease labels.
This direct approach can reduce transparency and potentially hinder safe clinical use. MedCORE aims to embed explicit clinical criteria into a vision-language system.
Core Components of the MedCORE Framework
For each image MedCORE decomposes the diagnostic process into clinically defined criteria. It spatially localizes each criterion to relevant regions encodes multi-scale evidence capturing macro-structural and micro-textural features and refines representations with a Graph Attention Network to model inter-criteria dependencies.
Criterion representations are aligned with clinical text descriptors reinforced by class-wise visual prototypes and aggregated via uncertainty-calibrated weighting that discounts low-confidence evidence.
Evaluation on Heterogeneous Datasets
The abstract describes validation across three imaging modalities: dermoscopic lesion classification on ISIC 2018 breast ultrasound lesion characterization on BUSI and diabetic retinopathy grading on IDRiD.
The preprint reports specific performance figures on these datasets and states that MedCORE shows gains relative to CNN transformer biomedical vision-language concept-based and prototype-based baselines.
Implications for Interpretable Medical AI
By grounding outputs in explicit criteria and spatial localization the framework may enhance interpretability. This could help clinicians better understand and audit model decisions.
The work highlights a potential bridge between black-box performance and the stepwise logic of medical practice. Because only the abstract is available the complete methodology results and limitations cannot yet be fully assessed.
