MSCRHO-Net:一种有效去除皮肤镜图像毛发遮挡的深度学习模型  

MSCRHO-Net:A Deep Learning Model for Effective Removal of Hair Occlusion from Dermoscopic Images

作　　者：杜红萱 刘庆一 任延德[2] 王艳 张亚楠 白培瑞 Du Hongxuan;Liu Qingyi;Ren Yande;Wang Yan;Zhang Yanan;Bai Peirui(College of Electronic and Information Engineering,Shandong University of Science and Technology,Qingdao 266590,Shandong,China;Department of Radiology,Affiliated Hospital of Qingdao University,Qingdao 265000,Shandong,China)

机构地区：[1]山东科技大学电子信息工程学院,山东青岛266590 [2]青岛大学附属医院放射科,山东青岛265000

出　　处：《中国生物医学工程学报》2024年第6期662-672,共11页Chinese Journal of Biomedical Engineering

基　　金：青岛市医药卫生科研计划项目(2021WJZD192)。

摘　　要：基于皮肤镜图像的自动分析对早期检测和诊断皮肤癌具有重要意义,而毛发遮挡对图像特征提取和皮肤病变诊断性能提出挑战。本研究提出一种融合拉普拉斯金字塔结构的多尺度级联深度学习模型(MSCRHO-Net),可以有效应对皮肤镜图像不同形式的毛发遮挡问题。首先,利用拉普拉斯金字塔获得图像空间不同尺度的关键特征。对每个尺度设计一个级联块,以由粗到精的方式逐步预测无毛发图像,实现高精度的毛发提取并更好地保留图像边缘细节;然后,构造感知损失与SSIM损失相结合的损失函数,提高图像细节恢复的质量,得到更加清晰的去毛发图像。就所提模型在合成数据集和真实数据集ISIC2017(训练图像共4750张,合成测试图像400张,真实测试图像223张)上进行了实验验证。结果表明,MSCRHO-Net无需大量的参数学习,能够有效去除皮肤镜图像中的毛发。SSIM和PSNR均值指标分别达到了0.9584和35.49,与其他传统毛发去除方法相比性能有显著性提升(P<0.05)。MSCRHO-Net对复杂毛发结构表现出高度适应性和鲁棒性,可以有效应对病变纹理破损、图像模糊等问题。It is of great significance to carry out early detection and diagnosis of skin cancer based on automatic analysis of dermoscopic images.However,hair occlusion poses a challenge to image feature extraction and skin lesion diagnosis.In this paper,a multi-scale cascade deep learning model(MSCRHO-Net)by integrating the Laplace pyramid was proposed.First,Laplace pyramid was employed to extract the key features of different scales in image space.A cascade block was designed for each scale channel to predict the hairless image by a coarse to fine scheme.High precision hair extraction and boundaries details retention were achieved through this operation.Then,a combined loss function including perceptual loss and SSIM loss was constructed,which was helpful to enhance details recovery and obtain more clear hair removal images.The performance of MSCRHO-Net was validated on synthetic dataset and real dataset ISIC2017(4750 training images,400 synthetic test images,and 223 real test images).The experimental results demonstrated that MSCRHO-Net was able to remove hair effectively without learning of huge parameters.The mean values of SSIM and PSNR reached 0.9584 and 35.49 respectively,which significantly improved the performance(P<0.05)compared with other traditional hair removal methods.MSCRHO-Net shows high adaptability and robustness to complex hair structure,and can deal with complicated scenarios such as damaged lesion texture and blurred image.

关 键 词：皮肤镜图像 毛发去除 拉普拉斯金字塔 感知损失 递归神经网络 

分 类 号：R318[医药卫生—生物医学工程]