Abstract:Endoscopic images provide a critical foundation for the screening and diagnosis of gastric cancer. However, the accuracy of traditional endoscopic examinations remains limited. To address this issue, multimodal fusion-based anomaly detection techniques have been applied to endoscopic image analysis. However, they still suffer from modality bias and the scarcity of paired data. To overcome these limitations, this study proposes an endoscopic anomaly detection model integrating multimodal enhanced fusion and multi-branch knowledge distillation. The framework incorporates a cross-masked attention cross-modal fusion module that explores latent inter-modal relationships through local feature reconstruction and cross-attention mechanisms. Furthermore, a multi-branch cross-modal distillation architecture is established, comprising a multimodal teacher network and two independent student branches. This design requires only the teacher network to be trained on paired data while enabling the student branches to learn without any paired data, thus significantly reducing dependency on paired samples and effectively mitigating modality bias. Additionally, a global cosine similarity loss is introduced to enhance consistency in multimodal feature representation. Extensive experiments on public real-world datasets demonstrate that the proposed method achieves competitive performance in multimodal endoscopic anomaly detection tasks. Code will be released at: https://github.com/LuoYifei-xs/CEMD.