Imbalanced Classiﬁcation in Medical Imaging via Regrouping Le Peng1 Yash Travadi2 Rui Zhang3 Ying Cui4 Ju Sun1

2025-05-08 0 0 374.38KB 5 页 10玖币

侵权投诉

Imbalanced Classiﬁcation in Medical Imaging

via Regrouping

Le Peng1, Yash Travadi2, Rui Zhang3, Ying Cui4, Ju Sun1

1Computer Science & Engineering, University of Minnesota, Twin Cities

2School of Statistics, University of Minnesota, Twin Cities

3Department of Surgery, University of Minnesota, Twin Cities

4Industrial and Systems Engineering, University of Minnesota, Twin Cities

{peng0347,trava029,zhan1386,yingcui,jusun}@umn.edu

Abstract

We propose performing imbalanced classiﬁcation by regrouping majority classes

into small classes so that we turn the problem into balanced multiclass classiﬁca-

tion. This new idea is dramatically different from popular loss reweighting and

class resampling methods. Our preliminary result on imbalanced medical image

classiﬁcation shows that this natural idea can substantially boost the classiﬁca-

tion performance as measured by average precision (approximately area-under-

the-precision-recall-curve, or AUPRC), which is more appropriate for evaluating

imbalanced classiﬁcation than other metrics such as balanced accuracy.

The shaky foundation of machine learning (ML) for medical imaging

Modern data classiﬁca-

tion is founded on accuracy (ACC) maximization, or equivalently error minimization:

min

f∈H

E(x,y)∼D1{y6=f(x)}=X

P(i=y)Ex|y=i1{y6=f(x)},(1)

where

is the hypothesis class,

is the data distribution, and

is the one-hot-encoded vector of

But accuracy can be misleading if the data are imbalanced across classes, e.g.,

D(y)

not uniform.

This intrinsic class imbalance is prevalent in medical imaging classiﬁcation (MIC), e.g., in binary

cases, the negative rate is expected to far exceed the positive rate in most diseases, and in multiclass

cases, the prevalence rates of different diseases are generally disparate. A slightly reﬁned notion is

balanced-accuracy (BA), which leads to the balanced-error (BE) minimization:

min

f∈H

| {y} | X

Ex|y=i1{y6=f(x)}.(2)

Figure 1: An example confusion table

for binary classiﬁcation, and the various

associated performance metrics.

From a quick example in Fig. 1, it is clear that BA can

be more responsive to recall performance than ACC, but

it still fails to capture low precision which together with

recall probably matters the most for medical diagnosis.

Hence, for MIC, reporting precision-recall and the as-

sociated F1 score seems most appropriate [

]. But

precision/recall and the F1 score depend on the decision

threshold, for which our typical choice following the bal-

anced scenarios is provably suboptimal with class imbal-

ance [

]. Hence, the

area under the precision-recall

curve

(AUPRC), which is often approximated by

average

precision

(AP), seems a more sensible choice, as it only

depends on the ordering of the prediction scores and not

Preprint. Under review.

arXiv:2210.12234v2 [cs.CV] 26 Oct 2022

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

ImbalancedClassicationinMedicalImagingviaRegroupingLePeng1,YashTravadi2,RuiZhang3,YingCui4,JuSun11ComputerScience&Engineering,UniversityofMinnesota,TwinCities2SchoolofStatistics,UniversityofMinnesota,TwinCities3DepartmentofSurgery,UniversityofMinnesota,TwinCities4IndustrialandSystemsEngineering,Uni...

展开>> 收起<<

Imbalanced Classiﬁcation in Medical Imaging via Regrouping Le Peng1 Yash Travadi2 Rui Zhang3 Ying Cui4 Ju Sun1.pdf

共5页,预览1页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Imbalanced Classiﬁcation in Medical Imaging via Regrouping Le Peng1 Yash Travadi2 Rui Zhang3 Ying Cui4 Ju Sun1

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: