上下文空间与实例信息的皮肤镜图像自监督分类

Dermoscopic images classification based on context and instance-level feature of self-supervised learning

针对基于深度学习的皮肤疾病检测方法受限于皮肤镜图像标注难度大、成本高、耗时耗力等问题,本文提出了一种针对皮肤镜图像的自监督学习方法,利用无标注的皮肤镜图像来提高皮肤疾病分类准确率。根据皮肤镜图像空间结构相似度高、信息单一的特点,使用基于旋转预测自监督代理任务,通过模型预测图像旋转的角度,约束模型对图像病变区域的关注,学习上下文空间特征;同时使用个体判别任务,利用正负样本对比学习,获得皮肤镜图像实例信息,为皮肤镜图像的分类提供全局分类信息指导;融合上下文空间特征与实例语义信息得到初始的皮肤疾病检测模型。实验结果表明:本文所提的自监督学习方法,从无标记的皮肤镜样本挖掘出重要信息,更有效地关注皮肤镜图像病变区域,通过融合皮肤镜图像特征信息,使分类结果更加准确。

The skin disease detection method based on deep learning is limited by the difficulty of annotations,high cost,time consuming and labor-intensive labeling of dermoscopy images.This paper proposes a self-supervised learning method for dermoscopy images,which uses unmarked dermoscopy images to improve the classification accuracy of skin diseases.According to the characteristics of high similarity and single information of the spatial structure of dermoscopic images,a rotation prediction pretext task is designed to predict the rotation angle of the image through the model,so as to constrain the model’s attention to pathological areas of the image,and learn the context feature.At the same time,the paper presents an instance discrimination task and comparative learning of positive and negative instances to obtain the instance information of dermatoscope images,which provides guidance for global classification information in the classification of dermatoscope images.The initial skin disease detection model is obtained by fusing context features and instance semantic information.The experimental results show that the self-supervised learning method proposed in this paper can mine important information from unmarked dermoscopy samples,pay more attention to the lesion area of dermoscopy images,and make the classification results more accurate by fusing the characteristic information of dermoscopy images.