CLINICAL Targeted Active Learning for Imbalanced Medical Image Classiﬁcation Suraj Kothawade1 Atharv Savarkar2 Venkat Iyer2 Lakshman Tamil1 Ganesh

2025-04-27 0 0 2.34MB 18 页 10玖币

侵权投诉

CLINICAL: Targeted Active Learning for

Imbalanced Medical Image Classiﬁcation

Suraj Kothawade

, Atharv Savarkar

, Venkat Iyer

, Lakshman Tamil

, Ganesh

Ramakrishnan2, and Rishabh Iyer1

1University of Texas at Dallas, USA

2Indian Institute of Technology, Bombay, India

suraj.kothawade@utdallas.edu

Abstract.

Training deep learning models on medical datasets that per-

form well for all classes is a challenging task. It is often the case that

a suboptimal performance is obtained on some classes due to the nat-

ural class imbalance issue that comes with medical data. An eﬀective

way to tackle this problem is by using targeted active learning, where

we iteratively add data points that belong to the rare classes, to the

training data. However, existing active learning methods are ineﬀective

in targeting rare classes in medical datasets. In this work, we propose

Clinical

(targeted a

tive

earning for

mbala

ced med

al im

assiﬁcation) a framework that uses submodular mutual information

functions as acquisition functions to mine critical data points from rare

classes. We apply our framework to a wide-array of medical imaging

datasets on a variety of real-world class imbalance scenarios - namely,

binary imbalance and long-tail imbalance. We show that

Clinical

outper-

forms the state-of-the-art active learning methods by acquiring a diverse

set of data points that belong to the rare classes.

1 Introduction

Owing to the advancement of deep learning, medical image classiﬁcation has made

tremendous advances in the past decade. However, medical datasets are naturally

imbalanced at the class level, i.e., some classes are comparatively rarer than the

others. For instance, cancerous classes are naturally rarer than non-cancerous ones.

In such scenarios, the over-represented classes overpower the training process and

the model ends up learning a biased representation. Deploying such biased models

results in incorrect predictions, which can be catastrophic and even lead to loss of

life. Active learning (AL) is a promising solution to mitigate this imbalance in the

training dataset. The goal of AL is to select data points from an unlabeled set for

addition to the training dataset at an additional labeling cost. The model is then

retrained with the new training set and the process is repeated. Reducing the

labeling cost using the AL paradigm is crucial in domains like medical imaging,

where labeling data requires expert supervision (e.g., doctors), which makes the

process extremely expensive. However, current AL methods are ineﬃcient in

selecting data points from the rare classes in medical image datasets. Broadly,

arXiv:2210.01520v1 [cs.CV] 4 Oct 2022

2 S. Kothawade et al.

they use acquisition functions that are either: i) based on the uncertainty scores

of the model, which are used to select the top uncertain data points [

], or ii)

based on diversity scores, where data points having diverse gradients are selected

[

]. They mainly focus on improving the overall performance of the model, and

thereby fail to target these rare yet critical classes. Unfortunately, this leads to a

wastage of expensive labeling resources when the goal is to improve performance

on these rare classes.

2330 2181

852 849

25810

5292

2443

873 708

Fig. 1: Motivating examples of two main

class imbalance scenarios occurring in

medical imaging.

Left:

Long-tail im-

balance (Diabetic retinopathy grading

from retinal images in APTOS-2019 [

]).

Right:

Binary imbalance (Microscopic

peripheral blood cell image classiﬁcation

in Blood-MNIST [

]). Red boxes in both

scenario denote targeted rare classes.

In this work, we consider two types

of class imbalance that recur in a

wide array of medical imaging datasets.

The ﬁrst scenario is binary imbal-

ance, where a subset of classes is

rare/infrequent and the remaining sub-

set is relatively frequent. The second

scenario is that of long-tail imbalance,

where the frequency of data points

from each class keeps steeply reducing

as we go from the most frequent class

to the rarest class (see Fig. 1). Such

class imbalance scenarios are particu-

larly challenging in the medical imag-

ing domain since there exist subtle dif-

ferences which are barely visually evi-

dent (see Fig. 1). In Sec. 3, we discuss

Clinical

, a targeted active learning

algorithm that acquires a subset by

maximizing the submodular mutual

information with a set of misclassiﬁed

data points from the rare classes. This

enables us to focus on data points from

the unlabeled set that are critical and

belong to the rare classes.

1.1 Related work

Uncertainty based Active Learning.

Uncertainty based methods aim to

select the most uncertain data points according to a model for labeling. The most

common techniques are - 1) Entropy [

] selects data points with maximum

entropy, 2) Least Confidence [

] selects data points with the lowest conﬁdence,

and 3) Margin [

] selects data points such that the diﬀerence between the top

two predictions is minimum.

Diversity based Active Learning.

The main drawback of uncertainty based

methods is that they lack diversity within the acquired subset. To mitigate this,

a number of approaches have proposed to incorporate diversity. The Coreset

method [

] minimizes a coreset loss to form coresets that represent the geometric

structure of the original dataset. They do so using a greedy k-center clustering. A

CLINICAL: Targeted AL for Imbalanced Medical Image Classiﬁcation 3

recent approach called Badge [

] uses the last linear layer gradients to represent

data points and runs K-means++ [

] to obtain centers that have a high gradient

magnitude. The centers being representative and having high gradient magnitude

ensures uncertainty and diversity at the same time. However, for batch AL,

Badge models diversity and uncertainty only within the batch and not across

all batches. Another method, BatchBald [

] requires a large number of Monte

Carlo dropout samples to obtain signiﬁcant mutual information which limits its

application to medical domains where data is scarce.

Class Imbalanced and Personalized Active Learning.

Closely related to

our method

Clinical

, are methods which optimize an objective that involves a

held-out set. GradMatch [

] uses an orthogonal matching pursuit algorithm

to select a subset whose gradient closely matches the gradient of a validation set.

Another method, Glister-Active [

] formulates an acquisition function that

maximizes the log-likelihood on a held-out validation set. We adopt GradMatch

and Glister-Active as baselines that targets rare classes in our class imbalance

setting and refer to it T-GradMatch and T-Glister in Sec. 4. Recently, [

]

proposed the use of submodular information measures for active learning in real-

istic scenarios, while [

] used them to ﬁnd rare objects in an autonomous driving

object detection dataset. Finally, [

] use the submodular mutual information

functions (used here) for personalized speech recognition. Our proposed method

uses the submodular mutual information to target selecting data points from the

rare classes via using a small set of misclassiﬁed data points as exemplars, which

makes our method applicable to binary as well as long-tail imbalance scenarios.

1.2 Our contributions

We summarize our contributions as follows:

We emphasize on the issue of

binary and long-tail class imbalance in medical datasets that leads to poor

performance on rare yet critical classes.

Given the limitations of current

AL methods on medical datasets, we propose

Clinical

, a novel AL framework

that can be applied to any class imbalance scenario.

We demonstrate the

eﬀectiveness of our framework for a diverse set of image classiﬁcation tasks

and modalities on Pneumonia-MNIST [

], Path-MNIST [

], Blood-MNIST

[

], APTOS-2019 [

], and ISIC-2018 [

] datasets. Furthermore, we show that

Clinical

outperforms the state-of-the-art AL methods by up to

≈

−

10% on

an average in terms of the average rare classes accuracy for binary imbalance

scenarios and long-tail imbalance scenarios.

We provide valuable insights about

the choice of submodular functions to be used for subset selection based on the

modality of medical data.

2 Preliminaries

Submodular Functions:

We let

denote the ground-set of

data points

{

, ..., n}

and a set function

: 2

V−→ R

. The function

is submodular [

] if

it satisﬁes diminishing returns, namely

(

j|X

)

≥f

(

j|Y

)for all

X ⊆ Y ⊆ V, j /∈ Y

Facility location, graph cut, log determinants, etc. are some examples [9].

4 S. Kothawade et al.

Submodular Mutual Information (Smi):

Given a set of items

A,Q ⊆ V

the submodular mutual information (MI) [6, 8] is deﬁned as If(A;Q) = f(A) +

(

)

−f

(

A ∪ Q

). Intuitively, this measures the similarity between

and

we refer to

as the query set. [

] extend Smi to handle the case when the target

can come from a diﬀerent set

apart from the ground set

. In the context of

imbalanced medical image classiﬁcation,

is the source set of images and the

query set

is the target set containing the rare class images. To ﬁnd an optimal

subset given a query set

Q⊆V0

, we can deﬁne

(

) =

(

;

A⊆V

and

maximize the same.

2.1 Examples of Smi functions

For targeted active learning, we use the recently introduced Smi functions in [

]

and their extensions introduced in [

] as acquisition functions. For any two data

points i∈ V and j∈ Q, let sij denote the similarity between them.

Graph Cut MI (Gcmi):

The Smi instantiation of graph-cut (Gcmi) is deﬁned

as:

IGC

(

;

)=2

Pi∈A Pj∈Q sij

. Since maximizing Gcmi maximizes the joint

pairwise sum with the query set, it will lead to a summary similar to the query

set

. In fact, speciﬁc instantiations of Gcmi have been intuitively used for

query-focused summarization for videos [28] and documents [21, 20].

Facility Location MI (Flmi):

We consider two variants of Flmi. The ﬁrst

variant is deﬁned over

(Flvmi), the Smi instantiation can be deﬁned as:

IF LV

(

;

) =

Pi∈V min

(

maxj∈A sij ,maxj∈Q sij

). The ﬁrst term in the min(.)

of Flvmi models diversity, and the second term models query relevance.

For the second variant, which is deﬁned over

(Flqmi), the Smi instantiation

can be deﬁned as:

IF LQ

(

;

) =

Pi∈Q maxj∈A sij

Pi∈A maxj∈Q sij

.Flqmi

is very intuitive for query relevance as well. It measures the representation of

data points that are the most relevant to the query set and vice versa.

Log Determinant MI (Logdetmi):

The Smi instantiation of Logdetmi can

be deﬁned as:

ILogDet

(

;

) =

log det

(

)

−log det

(

SA−SA,QS−1

QST

A,Q

SA,Q

denotes the cross-similarity matrix between the items in sets Aand Q.

3Clinical: Our Targeted Active Learning framework

for Binary and Long-tail Imbalance

In this section, we propose our targeted active learning framework,

Clinical

(see Fig. 2), and show how it can be applied to datasets with class imbalance.

Concretely, we apply the Smi functions as acquisition functions for improving a

model’s accuracy on rare classes at a given additional labeling cost (

instances)

without compromising on the overall accuracy. The main idea in

Clinical

, is

to use only the misclassiﬁed data points from a held-out target set

containing

data points from the rare classes. Let

T ⊆ T

be the subset of misclassiﬁed data

points. Then, we optimize the Smi function

(

;

)using a greedy strategy [

Note that since

contains only the misclassiﬁed data points, it would contain

more data points from classes that are comparatively rarer or the worst perform-

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：

CLINICAL:TargetedActiveLearningforImbalancedMedicalImageClassicationSurajKothawade1,AtharvSavarkar2,VenkatIyer2,LakshmanTamil1,GaneshRamakrishnan2,andRishabhIyer11UniversityofTexasatDallas,USA2IndianInstituteofTechnology,Bombay,Indiasuraj.kothawade@utdallas.eduAbstract.Trainingdeeplearningmodelsonm...

展开>> 收起<<

CLINICAL Targeted Active Learning for Imbalanced Medical Image Classiﬁcation Suraj Kothawade1 Atharv Savarkar2 Venkat Iyer2 Lakshman Tamil1 Ganesh.pdf

共18页,预览4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！

CLINICAL Targeted Active Learning for Imbalanced Medical Image Classiﬁcation Suraj Kothawade1 Atharv Savarkar2 Venkat Iyer2 Lakshman Tamil1 Ganesh

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: