皮肤镜图像多分类算法研究

Dermoscopy, as a non-invasive skin imaging technique which makes subsurface structure more easily visible, is an effective diagnostic tool for skin cancer as well as various skin diseases. Compared with skin cancer, the probability of having common skin diseases is much higher, therefore the research on multi-classification for common skin diseases is more clinically meaningful. . This project adopts deep learning method to study the multi-classification of 6 skin diseases. Firstly, a multi-scale U-net network that combines saliency information as a priori of lesion region is designed, to mines segmentation features. And then, to extract classification features, a convolution network combining nested residual structure with cyclic residual structure is designed and the loss function is improved, which increases the learning of feature diversity, and stable and effective classification features can be extracted. Finally, the multi-task learning framework is used to design an aided diagnosis model, and through the transfer learning, the segmentation features and classification features are deeply fused, and the multi-classification of dermoscopy images is achieved. . Furthermore, we extract the region with large deep feature response, and analyze the match relation between the lesion patterns in this region and the clinical diagnosis indications, which can provide reference for physicians in the correction and improvement of the clinical diagnosis indications. This project not only can help clinical physicians to make a right diagnosis, but also can help the general people to have the skin physical examination by themselves through internet.

皮肤镜能够观察活体皮肤表面以下微细结构和色素，是皮肤癌及各种皮肤疾病的有效诊断工具。相比于皮肤癌，常见皮肤疾病的患病概率更大，因此对常见皮肤疾病的多分类研究更具临床意义。本课题采用深度学习方法，研究6种皮肤疾病的多分类。设计多尺度U-net网络，结合显著性信息作为皮损目标区域先验，挖掘皮损目标与背景皮肤相区别的分割特征。设计嵌套残差和旋转残差相结合的卷积网络，改进loss函数，解决反向传播中梯度消失问题，加强特征多样性的学习，挖掘稳定有效的目标分类特征。采用多任务学习框架，设计分割和分类一体化的辅助诊断系统模型，通过迁移学习将分割特征和分类特征进行深度融合，实现皮肤镜图像的多分类。更进一步，对深度特征响应部位的皮损模式与临床诊断指征进行匹配分析，为医生对临床诊断指征的修正和完善提供参考。课题研究将辅助临床医师做出正确诊断，同时也为普通人群的皮肤健康自检提供帮助。

深度学习网络能够自适应提取出表征皮肤信息的特征，并将这些量化后的特征用于皮损的分割、分类和检索等任务，从而辅助医生进行皮肤疾病的临床诊断。本课题面向皮肤镜图像，基于深度学习理论，研究常见皮损的多分类方法。首先提出了一种基于高分辨率特征图的卷积结构，并设计了基于区域均值池化的分类模型，将分割框架与多分类框架融合，实现了分割和分类一体化的皮肤镜图像多分类。皮损具有旋转不变的特点，本课题提出了一种卷积神经网络的通用操作，即旋转均值操作，提高了卷积网络的抗旋转能力。更进一步，为了贴近皮肤病临床诊断的实际需求，本课题将临床图像和皮肤镜图像相融合，实现不同模态数据间的信息互补与整合，提升了皮损分类的准确率。此外，本课题进一步研究皮损图像的检索任务，基于残差DenseNet，提出了结合注意力机制和柯西旋转不变损失的皮肤镜图像检索方法，为医生提供历史病例的参考。本课题还对辅助诊断方法的临床应用进行思考，并设计实验方法研究智能模型在临床中的应用模式。本课题对以上研究进行了广泛的测试和分析，在主观和客观对比实验、人机对比、临床辅助诊断分析等实验中，全面验证了所提出方法的有效性。