皮肤镜图像质量的综合评价方法研究

Dermoscopy technology, through which the subsurface structures of the skin can be well revealed, is an effective tool to diagnose the skin cancer. And dermoscopy image quality influences the subsequent analysis result of skin tumors. Considering the typical factors which result in the low image qualities, including defocus blur, uneven illumination and hair noises, non-reference quality assessment method for dermoscopy images is researched. Firstly, the degrees of the defocus blur and the uneven illumination are evaluated by using the defocus radius as blur index and the gradient change of fitting brightness surface as uneven illumination index respectively. Secondly, the hair is extracted from images through connected component elongate and the hair noise degree is evaluated through the hair distribution in the image. Thirdly, the influences to the consequent image segmentation and diagnose result are analyzed when the three factors exist in the same image, and the comprehensive quality score is calculated through multi-factor comprehensive evaluation method. As the result of above research work, the quality standards for dermoscopy image are developed, and the images with low quality score are refused from the analysis system to ensure the segmentation accurateness and the diagnose rightness.

皮肤镜能够观察活体皮肤表面以下微细结构和色素，是皮肤癌早期发现和诊断的有效工具。皮肤镜图像的采集质量直接影响了后续图像分析诊断的结果。针对皮肤镜图像在采集过程中可能出现的散焦模糊、光照不均、毛发遮挡等影响图像质量的典型因素，研究皮肤镜图像的无参考质量评价方法。从失真原理入手，为散焦模糊和光照不均信号建模，研究散焦半径的精确估计，用散焦半径的大小来衡量散焦模糊的程度，研究光照曲面的建模方法，用光照曲面的梯度变化来估计光照不均的失真程度。提取毛发目标，分析毛发的分布特性，进而评价毛发对图像的遮挡程度。研究散焦模糊、光照不均、毛发遮挡等各种因素混合存在时对图像质量的综合影响，以后续图像分割和分析诊断的结果为驱动，建立图像质量综合评价模型，对多种质量干扰因素混合存在时的皮肤镜图像质量做出正确评价。制定皮肤镜图像的质量标准，确保不合格的图像能够重新采集，从而为皮肤镜图像分析系统的有效诊断提供保障。

皮肤镜能够观察活体皮肤表面以下微细结构和色素，是皮肤癌早期发现和诊断的有效工具。皮肤镜图像的采集质量直接影响了后续图像分析诊断的结果。针对皮肤镜图像在采集过程中可能出现的影响图像质量的典型因素，本课题研究了皮肤镜图像的无参考质量评价方法。针对多种失真因素同时存在的情况，以后续图像分割和目标分类结果为驱动，提出了基于模糊神经网络的综合质量评价算法。针对轻度失真的图像，提出了基于全卷积神经网络的光照不均去除方法和以及散焦模糊的盲复原方法，提高了皮肤镜图像的质量。本课题将质量评价与传统的皮肤镜图像分析系统相结合，设计了带质量评价功能的皮肤镜图像分析系统。在该框架下，课题组研究了基于显著性的皮肤镜图像分割方法，研究皮损目标的特征提取，设计组合神经网络，实现皮肤镜图像的有效分类。. 本课题共录用和发表论文31篇，其中SCI检索论文19篇，EI检索论文10篇，申请专利1项，培养博士和硕士生7名，超额完成了预期研究任务。