基于类激活图的弱监督皮肤镜图像分割方法

Weakly supervised dermoscopic image segmentation method based on class activation maps

皮肤镜图像中病灶区域的精确分割是实现皮肤病自动化检测的关键步骤。现存的皮肤镜图像分割方法主要基于全监督图像分割,这需要大量的像素标注,费时费力。针对此问题,提出一种基于类激活图(CAM)的弱监督皮肤镜图像分割方法。首先,对原始图像进行预处理,去除图像中的毛发并对图像进行颜色归一化处理;然后,结合图像的多尺度输入,并在显著图的引导下,通过特征提取网络得到图像的类激活图;之后,将得到的类激活图通过条件随机场得到伪掩膜;最后,使用伪掩膜训练分割网络。在ISIC2017数据集上评估所提方法,结果显示,所提方法生成的伪掩膜的Dice系数达到82.64%,相似性系数达到71.92%,灵敏度达到90.01%,表明所提方法能够在大量减少人工标注工作量的同时生成高质量的伪掩膜。

Accurate segmentation of lesion areas in dermoscopic images is a key step to achieve automated dermatology detection.The existing dermoscopic image segmentation methods are mainly based on fully supervised image segmentation,which requires a lot of pixel annotations and is time-consuming and laborious.To solve this problem,a weakly supervised dermoscopic image segmentation method based on Class Activation Map(CAM)was proposed.First,the original image was preprocessed to remove hairs from the image and to normalize the color of the image.Then,combined with the multi-scale input of the image and guided by the saliency map,the class activation maps of the image were obtained by the feature extraction network.After that,the class activation maps were passed through the conditional random field to obtain the pseudo mask.Finally,the segmentation network was trained using the pseudo mask.The proposed method was evaluated on the ISIC2017 dataset,and the results show that the proposed method generates pseudo mask with 82.64%Dice coefficient,71.92%Jaccard index,and 90.01%sensitivity,which indicates that the proposed method is able to generate high-quality pseudo mask while significantly reducing the labor of manual annotation.