大选择性核双边网络的长尾分布医学图像分类方法  

Long-tail Distributed Medical Image Classification Based on Large Selective Nuclear Bilateral-branch Networks

作　　者：孙汤慧 赵刚[1,2] 郭美倩 SUN Tanghui;ZHAO Gang;GUO Meiqian(School of Mathematics and Information Science,Nanchang Hangkong University,Nanchang 330063,China;Key Laboratory of Nondestructive Testing Technology of the Ministry of Education,Nanchang Hangkong University,Nanchang 330063,China)

机构地区：[1]南昌航空大学数学与信息科学学院,南昌330063 [2]南昌航空大学无损检测技术教育部重点实验室,南昌330063

出　　处：《计算机科学》2025年第4期231-239,共9页Computer Science

基　　金：国家自然科学基金(62366033);无损检测技术教育部重点实验室开放基金(EW202107216);江西省自然科学基金(20242BAB25121)。

摘　　要：医学场景下的数据集通常呈现长尾分布的特点,这种不平衡性可能导致模型偏向头部类,而对尾部类的识别性能较差,从而影响模型的准确性。常见的解决方法是对原始数据进行数据增强,使其具备平衡分布的特点,但增强后的尾部类样本质量往往不佳,没有真正改善尾部类的分类精度。针对此问题,提出一种大选择性核双边网络模型(LSKBB)。该模型主要由传统学习分支和重新再平衡分支两部分组成,采用LSK模块来获取关键信息和关注上下文信息,设计了可以使模型由一个关注方向逐渐过渡到另一个关注方向的动态损失函数,从而提高分类精度。在不改变长尾分布特点的医学数据集中进行图像分类实验,与现有方法相比,所提出的LSKBB模型性能在不平衡率为10,50和100时,在BreaKHis数据集下,准确率分别提高1.41%,1.25%和1.25%;在ChestX-ray数据集下,准确率分别提高6.10%,3.15%和2.47%。实验结果表明,LSKBB模型在不同的不平衡率下性能较好,可用于长尾分布的医学数据集的分类检测。In medical scenarios,datasets often exhibit characteristics of a long-tailed distribution,where in the imbalance may cause models to favor head classes,resulting in poorer performance in identifying tail classes and thus affecting model accuracy.Common approaches involve data augmentation to transform original data into a balanced distribution.However,the quality of augmented tail class samples is often inadequate,failing to genuinely improve the classification accuracy of tail classes.Addressing this issue,this paper proposes a large selective kernel bilateral branch network model(LSKBB).The model mainly consists of two parts:the traditional learning branch and the re-balancing branch.It adopts the LSK module to acquire key information and focus on contextual information.Additionally,a dynamic loss function is designed to enable the model to transition gradually from one focus direction to another,thereby enhancing classification accuracy.In image classification experiments conducted on medical datasets with long-tail distributions without altering their characteristics,the proposed LSKBB model shows performance improvements compared to existing methods.When the imbalance ratios are 10,50,and 100,the accuracy of the LSKBB model increases by 1.41%,1.25%,and 1.25%,respectively,on BreaKHis dataset.On ChestX-ray dataset,the accuracy increases by 6.10%,3.15%,and 2.47%,respectively.The experimental results indicate that the LSKBB model achieves good performance under different imbalance ratios and is suitable for classification and detection on medical datasets with long-tail distributions.

关 键 词：长尾分布 深度学习 双分支网络 LSK模块 图像分类 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]