四川大学学报（自然科学版）2025，Vol.62Issue(3)：524-536,13.DOI:10.19907/j.0490-6756.240366
基于交叉注意力的病灶区域感知网络用于多视角痤疮图像分级
Lesion region perception network with cross-attention for multi-view acne grading
摘要
Abstract
Automated grading of vulgaris acne can help dermatologists develop specific treatment plans for pa-tients.However,current research primarily focuses on single-view facial images,which may not capture com-plete facial information.Additionally,some methods lack sufficient attention to acne lesion areas,resulting in suboptimal grading accuracy.To address these problems,this paper proposes the Lesion Region Perception Network(LRPN)with Cross-Attention for Multi-View Acne Image Grading.It consists of the Multi-view Feature Fusion(MFF)module and the Lesion Region Perception(LRP)module.Specifically,the MFF module reduces redundancy between features and enhances complementary information by using cross-attention between different view features.Within the LRP module,acne lesion features in feature maps are identified through the multi-perception mechanism.This improvement enhances the model's focus on acne le-sions without requiring additional positional label information.Extensive experiments are conducted on the collected Multi-View Acne Grading dataset(MVACNE)and the publicly available datasets ACNE04 and ISIC2016.The experimental results show that the proposed method achieves good grading performance and effectively performs multi-view feature fusion and lesion area perception.关键词
痤疮分级/医学图像处理/注意力/多视角特征Key words
Acne grading/Medical image processing/Attention/Multi-view feature分类
信息技术与安全科学引用本文复制引用
张显良,张蕾,刘文杰,周新阳,李佳奇,杜丹,蒋献..基于交叉注意力的病灶区域感知网络用于多视角痤疮图像分级[J].四川大学学报（自然科学版）,2025,62(3):524-536,13.基金项目
国家杰出青年科学基金(62025601) （62025601）
