Random Data Augmentation based Enhancement A Generalized Enhancement Approach for Medical Datasets Sidra Aleem1 Teerath Kumar2 Suzanne Little2 Malika Bendechache2 Rob Brennan3 and Kevin

2025-04-29 0 0 7.22MB 9 页 10玖币

侵权投诉

Random Data Augmentation based Enhancement: A Generalized

Enhancement Approach for Medical Datasets

Sidra Aleem1, Teerath Kumar2, Suzanne Little2, Malika Bendechache2, Rob Brennan3, and Kevin

McGuinness1

1School of Electronic Engineering, Dublin City University, Ireland

2School of Computing, Dublin City University, Ireland

3ADAPT, School of Computer Science, University College Dublin, Ireland

Abstract

Over the years, the paradigm of medical image analysis has shifted from manual expertise to auto-

mated systems, often using deep learning (DL) systems. The performance of deep learning algorithms

is highly dependent on data quality. Particularly for the medical domain, it is an important aspect as

medical data is very sensitive to quality and poor quality can lead to misdiagnosis. To improve the di-

agnostic performance, research has been done both in complex DL architectures and in improving data

quality using dataset dependent static hyperparameters. However, the performance is still constrained due

to data quality and overﬁtting of hyperparameters to a speciﬁc dataset. To overcome these issues, this

paper proposes random data augmentation based enhancement. The main objective is to develop a gen-

eralized, data-independent and computationally efﬁcient enhancement approach to improve medical data

quality for DL. The quality is enhanced by improving the brightness and contrast of images. In contrast

to the existing methods, our method generates enhancement hyperparameters randomly within a deﬁned

range, which makes it robust and prevents overﬁtting to a speciﬁc dataset. To evaluate the generalization

of the proposed method, we use four medical datasets and compare its performance with state-of-the-art

methods for both classiﬁcation and segmentation tasks. For grayscale imagery, experiments have been

performed with: COVID-19 chest X-ray, KiTS19, and for RGB imagery with: LC25000 datasets. Experi-

mental results demonstrate that with the proposed enhancement methodology, DL architectures outperform

other existing methods. Our code is publicly available at: https://github.com/aleemsidra/

Augmentation-Based-Generalized-Enhancement.

Keywords: Classiﬁcation, Data Augmentation, Generalized Enhancement, Segmentation

1 Introduction

DL algorithms are being used in many domains such image data [Kumar et al., 2021a], audio data [Chandio et al., 2021,

Turab et al., 2022, Kumar et al., 2020, Park et al., 2020] and many more [Kumar et al., 2021b] and have revo-

lutionized the ﬁeld of medical image analysis. DL based applications are widely being used for computer

assisted disease diagnosis to aid clinicians [Philipp et al., 2021]. Regardless of the DL model architecture, per-

formance is strongly affected by the quality of the raw data [Zhou et al., 2019]. Particularly for medical images,

data quality is a critical factor for reliable disease assessment and diagnosis [Chen et al., 2014]. Further issues

arise due to the differences of acquisition protocols and the heterogeneity of data [Boyat and Joshi, 2015]. To

work around these issues, research has been done both to introduce new DL architectures [Szegedy et al., 2015,

Yadav and Jadhav, 2019, Cai et al., 2020] and to improve the data quality [Faes et al., 2019]. However, both

of these solutions pose further problems. The use of new complex DL architectures has a number of draw-

backs: (a) the issue of data quality remains persistent, which affects the ﬁnal prediction; (b) it does not

arXiv:2210.00824v1 [eess.IV] 3 Oct 2022

give an insight to the actual performance capability of traditional DL architectures; (c) it increases compu-

tational complexity without addressing the actual quality issue. To improve data quality, the optimal selection

of enhancement hyperparameters is extremely important. Due to the sensitive nature of medical data, this

selection is even more important to have a robust performance independent of the data quality. The exist-

ing methods rely on ﬁxed enhancement hyperparameters. These hyperparameters are chosen according to

the dataset [Masud et al., 2021, Mittal et al., 2019, Hari et al., 2013]. Consequently, such enhancement meth-

ods are prone to overﬁtting. Moreover, these techniques are evaluated either on grayscale or RGB datasets

[Gao et al., 2021, Wang et al., 2019, Zhou et al., 2019]. Thus, this the limited evaluation is not indicative of

the strength of such methods. To prevent these issues, a computationally efﬁcient and generalized contrast

enhancement method can help.

To overcome above stated issues, this paper proposes random data augmentation based computationally

efﬁcient and generalized enhancement method, in which data quality is improved using random brightness and

contrast hyperparameters. Unlike other existing methods, which use ﬁxed hyperparameters for enhancement,

we use a set of hyperparameters. The hyperparameters are randomly chosen from this set and hence are not

reliant on the dataset. The random selection ensures that the hyperparameters do not overﬁt the data. The data-

independent nature of these hyperparameters aids the proposed method to generalize well on different datasets.

The range of brightness set is [1.15, 1.35] and the range of contrast set is [−0.1, 0.4]. These speciﬁc ranges are

chosen by performing experimental evaluation. First we started from -1.0 for both hyperparameters, where

images have apparently no features. We evaluated the corresponding affect on enhancement by visualizing

resultant data and kept on incrementing the value with interval of 0.15. It was observed that images started to

show some feature at 1.15 and -0.1 brightness and contrast values respectively. Thus, these values were chosen

as starting points for contrast and brightness sets. For the end point, we followed the same methodology and

choose those values as end points before which images started to loose the features. The enhancement results

achieved with starting and end points of selected range are shown in Figure 4. Beyond this particular range,

image becomes darker or brighter and starts losing features. The performance is assessed by evaluating the

resultant enhanced data with a variety of traditional DL architectures.

Contributions: The main contributions of our work are as follows:

• We propose a generalized and computationally efﬁcient random data augmentation based enhancement

approach for medical data.

• The enhancement hyperparameters are not manually selected according to the data; thus our enhancement

method does not overﬁt a speciﬁc dataset.

• To check the effectiveness of our work, we perform extensive experiments on both gray scale and RGB

datasets for classiﬁcation and segmentation tasks.

• The proposed approach shows superior performance in terms of both accuracy and execution time over

state-of-the-art techniques.

2 Related Work

Generally three types of contrast enhancement methods have been used: histogram methods, spectral methods

and spatial methods [Pierre et al., 2017]. The histogram methods have remained very popular for contrast en-

hancement. Such methods transform gray scale images to an image with a speciﬁed histogram. However, such

methods result in poor enhancement that can be attributed to both loss of information and over-enhancement

of speciﬁc gray levels. Such methods are not adaptive and thus are inappropriate to provide contrast enhance-

ment for the medical imaging domain [Reddy et al., 2018, Singh et al., 2016]. Spectral methods use wavelet

transforms for quality enhancement. However, such methods fail to provide simultaneous enhancement to all

the parts of of images. Moreover, it is difﬁcult to automate enhancement using them [Wang et al., 2019]. The

motive of the contrast enhancement in medical images is to aid clinicians with automated diagnosis, so such

methods are also not best suited for medical domain.

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：

RandomDataAugmentationbasedEnhancement:AGeneralizedEnhancementApproachforMedicalDatasetsSidraAleem1,TeerathKumar2,SuzanneLittle2,MalikaBendechache2,RobBrennan3,andKevinMcGuinness11SchoolofElectronicEngineering,DublinCityUniversity,Ireland2SchoolofComputing,DublinCityUniversity,Ireland3ADAPT,Schoolof...

展开>> 收起<<

Random Data Augmentation based Enhancement A Generalized Enhancement Approach for Medical Datasets Sidra Aleem1 Teerath Kumar2 Suzanne Little2 Malika Bendechache2 Rob Brennan3 and Kevin.pdf

共9页,预览2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！

Random Data Augmentation based Enhancement A Generalized Enhancement Approach for Medical Datasets Sidra Aleem1 Teerath Kumar2 Suzanne Little2 Malika Bendechache2 Rob Brennan3 and Kevin

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: