Neurocomputingに当研究室の論文が採録

■書誌情報
Yun Jiang, Yarong Jin, Tao Sun, Huanting Guo, Zequn Zhang, Yuhang Li, Qian Niu: Enhancing Dermoscopic Image Generation via Multi-Modal Conditional Diffusion Models, Neurocomputing, 2026
■概要
Recent advancements in diffusion models have shown promise in alleviating data scarcity for dermoscopic image generation. However, existing methods often lack precise control over critical details, such as disease patterns and lesion localization. To address this issue, we introduce a novel Multimodal Conditional Diffusion Model (MMC-Diff), which leverages multiple modalities (e.g., dermatology-specific texts and images) to condition the generation process. First, we develop a latent diffusion model conditioned on dermatology-specific textual descriptions derived from a pre-trained vision-language model. Second, we present a cost-effective spatial mask adapter that aligns the model’s internal knowledge with external control signals, enabling accurate spatial control during image generation. Experimental results indicate that our model achieves competitive generation quality compared to state-of-the-art methods, based on metrics such as FID and IS. MMC-Diff effectively enhances the quality and diversity of dermoscopic datasets.