多尺度与三维交互特征优化的皮肤镜图像分类  

Dermoscopic image classification based on multi-scale and three-dimensional interaction feature optimization

作　　者：王迪 吕晓琪[1,2] 李菁 WANG Di;LÜXiaoqi;LI Jing(School of Digital and Intelligence Industry,Inner Mongolia University of Science and Technology,Baotou 014010,China;School of Information Engineering,Inner Mongolia University of Technology,Hohhot 010051,China)

机构地区：[1]内蒙古科技大学数智产业学院,内蒙古包头市014010 [2]内蒙古工业大学信息工程学院,内蒙古呼和浩特010051

出　　处：《光学精密工程》2024年第24期3644-3657,共14页Optics and Precision Engineering

基　　金：国家自然科学基金项目(No.61771266)。

摘　　要：皮肤癌的早期诊断有助于增加病患的治愈机会,减轻医疗系统的负担。针对皮肤镜图像分类存在的特征提取过程中会造成信息丢失与难以同时识别图像中独立的不同类型特征信息等问题,提出了一种多尺度与三维交互特征优化网络(MTIFNet)。首先,该网络应用多尺度空间自适应模块,在训练过程中提取图像全局与局部上下文信息,利用多尺度增强病灶周围模糊的像素点与像素点间的联系。其次,设计了一个在不同维度之间建立联系、对信息进行交流和整合的三维交互特征优化模块。最后,使用交叉熵损失衡量预测概率分布与真实类别分布之间的差异,以优化模型的准确性。基于ISIC 2018与ISIC 2017数据集的实验结果表明,该网络分别将准确率、精确率、召回率、特异性分别提升至94.32%,91.61%,93.00%,98.39%和98.57%,98.20%,98.47%,99.13%。与目前流行的ResNeSt,ConvNext,Fcanet等分类网络相比,MTIFNet具备更好的特征提取与交互能力,有助于医生做出更准确的诊断和治疗决策。Early diagnosis of skin cancer is crucial for improving patient outcomes and alleviating the bur⁃den on the healthcare system.However,the process of feature extraction in skin cancer image classifica⁃tion often results in information loss and challenges in simultaneously identifying independent types of fea⁃tures in the images.MTIFNet was proposed,which was a network that integrates three-dimensional spa⁃tial attention with information fusion.Initially,the network employed a multi-scale spatial adaptive mod⁃ule to extract both global and local contextual information from images during training.This module en⁃hanced the connection between blurred pixels around lesions and the relationship in pixels at different scales.Subsequently,a three-dimensional interaction feature optimization module was introduced to facili⁃tate connections across different dimensions,enabling the exchange and integration of information.Final⁃ly,cross-entropy loss was used to measure the difference between the predicted probability distribution and the true class distribution to optimize the accuracy of the model.The experimental results based on the ISIC 2018 and ISIC 2017 datasets indicate that the network has improved accuracy,precision,recall,and specificity to 94.32%,91.61%,93.00%,98.39%and 98.57%,98.20%,98.47%,99.13%,respec⁃tively.Compared to currently popular classification networks such as ResNeSt,ConvNext,and Fcanet,MTIFNet demonstrates superior capabilities in feature extraction and interaction,thereby assisting health⁃care professionals in making more precise diagnostic and treatment decisions.

关 键 词：皮肤癌 图像分类 多尺度 空间自适应 三维交互 

分 类 号：TP394.1[自动化与计算机技术—计算机应用技术] TH691.9[自动化与计算机技术—计算机科学与技术]