基于PiT的皮肤镜图像分类方法研究  

Study on dermoscope image classification method based on PiT

作　　者：韦春苗 徐岩[1] 蒋新辉 魏一铭 WEI Chunmiao;XU Yan;JIANG Xinhui;WEI Yiming(School of Electronic and Information Engineering,Lanzhou Jiaotong University,Lanzhou,Gansu 730070,China;College of Artificial Intelligence,Liuzhou Railway Vocational and Technical College,Liuzhou,Guangxi 545616,China)

机构地区：[1]兰州交通大学电子与信息工程学院 [2]柳州铁道职业技术学院人工智能学院

出　　处：《光电子．激光》2022年第5期505-512,共8页Journal of Optoelectronics·Laser

摘　　要：随着计算机技术的进步,现有的Transformer被扩展成处理计算机视觉任务的网络结构。为提高黑色素瘤的早期确诊率以提高皮肤病患者的治愈率,本文提出一种改进的基于PiT(pyramid pooling transformer)的网络模型来实现对7种皮肤病变的皮肤镜图像进行自动分类。本文模型主要由PiT模块和抗干扰模块等2个部分组成,Pit继承了ViT的优点,并通过池化进行空间尺寸转换来提高模型的鲁棒性,经过预训练的PiT网络拥有大量的自然图像特征,且PiT部分网络可为下游的分类任务提供所需的图像特征,本文设计出抗干扰模块,用来抵抗皮肤镜图像中的干扰因素(如毛发、异物遮挡)的影响,从而提高模型性能、提高分类精度。实验结果表明,本文模型在ISIC 2018验证集上的分类准确率、精确率、召回率、F1-score值分别高达91.58%、83.59%、89.92%、86.34%,每秒传输帧数(frames per second,FPS)达到85 Hz与现有的几种先进的分类网络相比,分类性能和模型效率都有所提高,具有相对优势,证明本文模型具有一定的实用价值。With the advancement of computer technology,the existing Transformer has been expanded into a network structure for processing computer vision tasks.In order to improve the early diagnosis rate of melanoma and the cure rate of skin disease patients,this paper proposes an improved network model based on PiT(pyramid pooling transformer) to realize automatic classification of dermoscopic images of seven skin lesions.The model of this paper is mainly composed of the PiT module and the anti-interference module.Pit inherits the advantages of ViT and uses pooling to perform spatial size conversion to improve the robustness of the model.The pre-trained PiT network has a large number of natural image features,and the PiT part of the network can provide the required image features for downstream classification tasks.In this paper,an anti-interference module is designed to resist the influence of interference factors(such as hair and foreign object occlusion) in the dermoscopic image,thereby improving the performance of the model.Improve classification accuracy.Experimental results show that the classification accuracy,precision,recall,and F1-score values of this model on the ISIC 2018 verification set are as high as 91.58%,83.59%,89.92%,86.34%,and the number of frames per second(FPS) reaches 85 Hz.Compared with several existing advanced classification networks,the classification performance and model efficiency have been improved,and it has relative advantages,which proves that the model in this paper has certain practical value.

关 键 词：图像处理 图像分类 PIT 抗干扰模块 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]