Department of Mechanical Engineering, Massachusetts Institute of Technology, 77 Massachusetts Avenue, Cambridge, MA, 02139, USA.
Comput Biol Med. 2017 Oct 1;89:314-324. doi: 10.1016/j.compbiomed.2017.08.020. Epub 2017 Aug 24.
Non-invasive imaging techniques allow the monitoring of skin structure and diagnosis of skin diseases in clinical applications. However, hair in skin images hampers the imaging and classification of the skin structure of interest. Although many hair segmentation methods have been proposed for digital hair removal, a major challenge in hair segmentation remains in detecting hairs that are thin, overlapping, of similar contrast or color to underlying skin, or overlaid on highly-textured skin structure.
To solve the problem, we present an automatic hair segmentation method that uses edge density (ED) and mean branch length (MBL) to measure hair. First, hair is detected by the integration of top-hat transform and modified second-order Gaussian filter. Second, we employ a robust adaptive threshold of ED and MBL to generate a hair mask. Third, the hair mask is refined by k-NN classification of hair and skin pixels.
The proposed algorithm was tested using two datasets of healthy skin images and lesion images respectively. These datasets were taken from different imaging platforms in various illumination levels and varying skin colors. We compared the hair detection and segmentation results from our algorithm and six other hair segmentation methods of state of the art. Our method exhibits high value of sensitivity: 75% and specificity: 95%, which indicates significantly higher accuracy and better balance between true positive and false positive detection than the other methods.
非侵入性成像技术可用于临床应用中的皮肤结构监测和皮肤病诊断。然而，皮肤图像中的毛发会妨碍对感兴趣的皮肤结构的成像和分类。尽管已经提出了许多用于数字除毛的毛发分割方法，但毛发分割的一个主要挑战仍然是检测到细小、重叠、与底层皮肤对比度或颜色相似或覆盖在高度纹理化的皮肤结构上的毛发。
为了解决这个问题，我们提出了一种自动毛发分割方法，该方法使用边缘密度 (ED) 和平均分支长度 (MBL) 来测量毛发。首先，通过顶帽变换和改进的二阶高斯滤波器的集成来检测毛发。其次，我们采用 ED 和 MBL 的稳健自适应阈值来生成毛发掩模。最后，通过毛发和皮肤像素的 k-NN 分类来细化毛发掩模。
该算法分别在两组健康皮肤图像和病变图像数据集上进行了测试。这些数据集来自不同成像平台，在不同的光照水平和不同的肤色下采集。我们比较了我们的算法和其他六种最新毛发分割方法的毛发检测和分割结果。我们的方法具有较高的灵敏度：75%和特异性：95%，这表明与其他方法相比，我们的方法具有更高的准确性和更好的真阳性和假阳性检测之间的平衡。
