Abstract
Medical image classification aims to categorise clinically significant imaging patterns, thereby facilitating accurate and timely diagnosis. However, existing approaches predominantly rely on visual features extracted from raw pixel data, often overlooking fine-grained diagnostic cues grounded in medical expertise. To address this limitation, we propose a novel knowledge-enhanced Model, called KEM, that leverages medical large vision-language models (Medical LVLMs) as domain experts to generate descriptive priors, which are used to guide and support clinical decision-making. Specifically, we prompt Medical LVLMs to generate rich, Multi-Dimension clinical descriptions tailored to each input image, capturing nuanced semantics. These descriptive priors are then encoded and fused with visual features through a dual cross-attention module, which enables bidirectional interaction and alignment between modalities. This design allows the model to dynamically attend to both textual and visual cues, thereby enhancing its ability to recognise subtle disease patterns. Comprehensive experiments on four benchmark datasets demonstrate that the proposed method significantly outperforms state-of-the-art vision-only models and exhibits strong generalization across varied clinical settings.