Automatic Diagnosis of Myocarditis Disease in Cardiac MRI Modality using Deep Transformers and Explainable Artificial Intelligence Mahboobeh Jafari12 Afshin Shoeibi12 Navid Ghassemi1 Jonathan Heras3 Sai Ho Ling4 Amin

2025-05-02 0 0 1.76MB 27 页 10玖币

侵权投诉

Automatic Diagnosis of Myocarditis Disease in Cardiac MRI Modality using Deep

Transformers and Explainable Artificial Intelligence

Mahboobeh Jafari1,2,*, Afshin Shoeibi1,2, Navid Ghassemi1, Jonathan Heras3, Sai Ho Ling4, Amin

Beheshti5, Yu-Dong Zhang6, Shui-Hua Wang6, Roohallah Alizadehsani7, Juan M. Gorriz2,8,

U. Rajendra Acharya9, Hamid Alinejad Rokny10,11,12

1 Internship in BioMedical Machine Learning Lab, The Graduate School of Biomedical Engineering, UNSW

Sydney, Sydney, NSW, 2052, Australia.

2 Data Science and Computational Intelligence Institute, University of Granada, Spain.

3 Department of Mathematics and Computer Science, University of La Rioja, La Rioja, Spain.

4 Faculty of Engineering and IT, University of Technology Sydney (UTS), Australia.

5 Data Analytics Lab, Department of Computing, Macquarie University, Sydney, NSW 2109, Australia.

6 School of Computing and Mathematical Sciences, University of Leicester, Leicester, UK.

7 Intelligent for Systems Research and Innovation (IISRI), Deakin University, Victoria 3217, Australia.

8 Department of Psychiatry, University of Cambridge, UK.

9 School of Mathematics, Physics and Computing, University of Southern Queensland, Springfield, Australia.

10 BioMedical Machine Learning Lab, The Graduate School of Biomedical Engineering, UNSW Sydney,

Sydney, NSW, 2052, Australia.

.UNSW Data Science Hub, The University of New South Wales, Sydney, NSW, 2052, Australia

12 Health Data Analytics Program, AI-enabled Processes (AIP) Research Centre, Macquarie University, Sydney,

2109, Australia.

* Corresponding author: Afshin Shoeibi (Afshin.shoeibi@gmail.com)

Abstract

Myocarditis is a significant cardiovascular disease (CVD) that poses a threat to the health of many

individuals by causing damage to the myocardium. The occurrence of microbes and viruses, including

the likes of HIV, plays a crucial role in the development of myocarditis disease (MCD). The images

produced during cardiac magnetic resonance imaging (CMRI) scans are low contrast, which can make

it challenging to diagnose cardiovascular diseases. In other hand, checking numerous CMRI slices for

each CVD patient can be a challenging task for medical doctors. To overcome the existing challenges,

researchers have suggested the use of artificial intelligence (AI)-based computer-aided diagnosis

systems (CADS). The presented paper outlines a CADS for the detection of MCD from CMR images,

utilizing deep learning (DL) methods. The proposed CADS consists of several steps, including dataset,

preprocessing, feature extraction, classification, and post-processing. First, the Z-Alizadeh dataset was

selected for the experiments. Subsequently, the CMR images underwent various preprocessing steps,

including denoising, resizing, as well as data augmentation (DA) via CutMix and MixUp techniques.

In the following, the most current deep pre-trained and transformer models are used for feature

extraction and classification on the CMR images. The findings of our study reveal that transformer

models exhibit superior performance in detecting MCD as opposed to pre-trained architectures. In terms

of DL architectures, the Turbulence Neural Transformer (TNT) model exhibited impressive accuracy,

reaching 99.73% utilizing a 10-fold cross-validation approach. Additionally, to pinpoint areas of

suspicion for MCD in CMRI images, the Explainable-based Grad Cam method was employed.

Keywords: Myocarditis, Diagnosis, Cardiac MRI, Deep Learning, Transformers, Grad CAM

1. Introduction

The prevalence of cardiovascular diseases (CVDs) is now a major global public health issue [1].

According to the World Health Organization (WHO), CVDs currently rank among the leading causes

of mortality globally [2-3]. The heart is responsible for the circulation of blood, carrying oxygen and

nutrients throughout the body, and expelling carbon dioxide [4]. The most common CVDs include

coronary artery disease (CAD) [5], arrhythmia [6], cardiomyopathy [7], heart failure [8], congenital

heart disease [9], mitral regurgitation [10], angina [11], and myocarditis disease (MCD) [12]. Symptoms

of CVDs usually manifest in chest pain, arm and left shoulder pain, shortness of breath, nausea and

fatigue, cold sweats, headache, and dizziness [13-15]. Risk factors for CVDs include hypertension,

smoking, high cholesterol levels, diabetes, obesity, family history, and advanced age [16-18]. Clinical

studies have demonstrated that minimizing these risk factors significantly reduces the likelihood of

CVDs occurrence and ensures good health [1-4].

In recent years, the incidence of MCD has increased worldwide [19-22]. MCD is caused by

inflammation of the myocardium, which impairs the heart's ability to pump blood throughout the body,

leading to significant health issues such as arrhythmia, chest pain, and dyspnea in patients [19-20]. In

some cases, myocardial dysfunction can also result in blood clots, heart attack, stroke, heart injury,

heart failure, or even death [21-22]. Clinical studies have revealed that SARS-CoV-2, adenovirus, and

HIV are the primary causes of MCD [19-20]. While MCD may be asymptomatic in some cases, it often

causes chest pain, heart failure, fever, palpitations, fatigue, and even sudden death [21]. Medical

professionals use various screening methods to diagnose CVDs, including ECG, Echo, cardiac exercise

testing, CT, CMRI, and Holter monitoring [23-28]. Among these methods, CMRI imaging is considered

one of the most effective non-invasive methods for diagnosing CVDs, including MCD. Specialists use

CMRI images to analyze different heart regions, such as ventricular wall thickness and left ventricular

end-systolic volume [29-31]. CMRI imaging provides early diagnosis of cardiac dysfunction, enabling

rapid and reliable diagnosis of CVDs, including MCD [29-31]. Despite the advantages of CMRI,

analyzing medical images is a challenging task for doctors. For accurate diagnosis of MCD multi-slice

CMRI images must be acquired for each patient, which is typically a difficult task even for experienced

doctors [20-21]. Additionally, CMRI images may have low resolution and low contrast, making it

challenging to diagnose MCD [21]. Furthermore, CMRI images may show various artifacts that

complicate the diagnosis of MCDs by specialists.

To address these challenges, researchers have used Artificial Intelligence (AI) techniques to diagnose

MCD from CMR images [29-31]. AI techniques are generally categorized into ML and DL methods

[95-96]. In references [97-98], research on the diagnosis of CVDs from CMRI data using ML techniques

is reported. In this research, ML-based CADS involves dataset, pre-processing, feature extraction,

feature selection and classification stages [97-98]. The researchers of these papers have demonstrated

that the feature extraction is the most important part in ML-based CADS for the diagnosis of CVDs

[97-98]. As such, subsequent to the phases of data registration and preprocessing, researchers attempt

to improve the diagnosis of CVDs by combining different feature extraction methods. However,

combining multiple features to achieve high diagnosis accuracy requires extensive knowledge of

researchers in the field of ML [97]. To address this challenge, DL techniques have recently been

proposed for detecting or predicting CVDs such as MCD [99-100]. Unlike ML methods, DL techniques

employ feature engineering in an unsupervised manner. References [99-100] have reviewed various

papers on the diagnosis of CVDs from CMR images using DL techniques. These studies demonstrate

that researchers using DL techniques have achieved significant results in the classification and

segmentation of CMR images for the diagnosis of CVDs, including MCD [99-100].

In recent years, studies on the diagnosis of MCD from CMRI images using DL techniques have been

published [19-21]. Sharifrazi et al. [19] proposed the CNN-KCL model for MCD detection based on

CMRI images, and experiments were performed on the Z-Alizadeh dataset. The study aimed to combine

a 2D-CNN model with the k-means clustering method. Their findings showed an accuracy of 97.41%.

In another study by Shoeibi et al. [20], the cycle-GAN method was utilized with various pre-trained

models to diagnose MCD. The Z-Alizadeh dataset was also used in this study to implement the proposed

model. The cycle-GAN architecture was employed in the preprocessing step to develop synthetic CMRI

images, which were then applied to various pre-trained models. Among them, the EfficientNet V2

method achieved an accuracy of 99.33%. Moravvej et al. [21] introduced deep reinforcement learning

(RL) for MCD detection using CMRI images and presented an RLMD-PA method to diagnose

myocarditis. Lastly, several optimization methods were evaluated to enhance the accuracy and

efficiency of MCD diagnosis.

In this study, we propose a deep learning-based computer-aided diagnosis system (CADS) for MCD

diagnosis from CMRI images. The presented paper introduces various novelties in the pre-processing,

classification and post-processing sections. In the first novelty, we utilized CutMix and MixUp

techniques in the pre-processing section [32-33]. Notably, this method has not been previously

employed in other studies for to generate artificial CMR images to improve the accuracy of CVDs

diagnosis [97-100]. Additionally, transformers architectures are a new class of DL techniques that have

shown high performance in various medical imaging and signal processing applications [101-102].

Compared to other DL architectures, transformers techniques show high performance on low input

images [101-102]. As a second novelty, we have simulated and compared the latest transformers

architectures. Additionally, we compared these networks with the most recent pretrained models to

demonstrate the efficiency of our proposed method. We thoroughly reviewed various papers in the field

of CVDs diagnosis, including myocarditis, and found that transformers methods have not been used by

researchers [97-100]. The results demonstrated that we achieved the highest accuracy in diagnosing

MCD from CMR images by using the combination of CutMix and MixUp methods and some

transformers models. Another novelty is the use of an explainable artificial intelligence technique as

Grad-Cam, which is important for clinical diagnoses [103]. XAI is a rapidly growing field in medicine,

and current research is focused on utilizing these methods to diagnose various diseases from medical

images [104-105]. In this study, we employed the Grad-Cam technique as an XAI method, which

provides a detailed understanding of the performance of DL architectures used in the diagnosis of

myocarditis from CMR images. These assist specialist physicians in accurately diagnosing myocarditis

from CMR images.

Some of the authors of this paper have previously published a comprehensive review paper on the

diagnosis of cardiac diseases from CMR images using DL techniques [106]. This review paper

highlights that a limited number of researchers have employed XAI techniques in CVDs diagnosis.

Moreover, the reference [106] and other review papers in this field [97-100] indicate that the proposed

method, including CutMix and MixUp methods, transformer models and XAI, has not been previously

employed for the diagnosis of myocarditis from CMR images. The rest of the paper is categorized as

follow: Section 2 describes the proposed CADS in detail, including the dataset, preprocessing, and DL

model. Section 3 represents the evaluation parameters of the proposed method. The experiment results

with details are presented in Section 4. Section 5 is about post-processing. In section 6, we discussed

about limitation of study. The discussion section of the paper is presented in section 7. Finally, the

conclusion and future directions are provided in Section 8.

Fig. 1. Block diagram of the proposed method for diagnosis of MCD

2. Material and Methods

This section proposes a CADS to diagnose MCD using CMRI images. Figure (1) shows the block

diagram of the proposed method which includes dataset, preprocessing, DL model, and post-processing.

First, a DL model is implemented using the Z-Alizadeh dataset [19]. The Z-Alizadeh dataset consists

of 12,000 CMRI images of normal and MCD patients admitted to Shahid Rajaee Hospital, Tehran. In

the pre-processing step, denoising, resizing, and a new DA method to generate synthetic CMRI images

were performed. The CMRI images were denoised in this section and then resized to 224*224. In the

following, synthetic CMRI data was generated using a new DA model based on the CutMix [32] and

MixUp [33] approaches. The proposed DA method is used for the first time in MCD detection and is

the first novelty in this study. In the third step, state-of-the-art pre-trained models and transformers were

used for feature extraction and classification of CMR images. Pre-trained models included EfficientNet

B3 [35], EfficientNet V2 [36], HrNet [37], Inception [38], ResNetrs50 [39], ResNest50d [40], and

ResNet 50d [41]. In addition, the transformer models also included Beit [42], Cait [43], Coat [44], Deit

[45], Pit [46], Swin [47], TNT [48], Visformer [49], and ViT [50]. The transformer model for

diagnosing MCD is another novelty of this work. Finally, the Grad-Cam technique, an explainable AI

technique, was used to visualize the suspicious regions of MCD in CMRI images [34].

2.1. Z-Alizadeh Dataset

The Z-Alizadeh Sani myocarditis dataset was collected between September 2018 and September 2019

at the CMR department of OMID hospital in Tehran, Iran. The soundness of the data gathering process

has been confirmed by the local ethical committee of OMID hospital. To perform CMR examination, a

1.5-T system (MAGNETOM Aera Siemens, Erlangen Germany) was used. Dedicated body coils were

used to scan each patient in the standard supine position. The CMR protocols that have been complied

with are listed below:

* CINE-segmented images and pre-contrast T2-weighted (trim) images were performed in short and

long axes views.

* The pre-contrast T1-weighted relative images were acquired in axial views of the myocardium.

* After injection of Gadolinium ((DOTAREM 0/1 mmol/kg), the T1-weighted relative sequence was

repeated. After 10-15 minutes, sequences of Late Gadolinium Enhancements (LGE- high-resolution

PSIR) in short and long axes views were carried out.

The total number of images is 10425. The number of images representing HC and MCD patients were

5040 and 5385, respectively. Figure (2) shows typical CMRI images obtained from the Z-Alizadeh

dataset for healthy control (HC) and MCD patients.

Fig 2. Sample CMR images dataset: a) Normal, and B) Abnormal.

2.2. Preprocessing

This section describes preprocessing steps used for CMRI images, involving denoising, image resizing,

and DA. Preprocessing is a crucial step in the analysis of medical images, particularly MRI data. In this

work, pre-processing steps entail noise removal, image size reduction and DA. CMR images contain

various artifacts during registration, posing a challenge for specialist physicians to accurately diagnose

myocarditis. Additionally, some slices of CMR images may have low contrast for certain subjects. To

diagnose myocarditis, specialist doctors record CMR images of different regions of the heart, resulting

in a dataset of images with varying sizes. To address this issue, the next preprocessing step involves

resizing all images to 224x224 dimensions. This will allow uniform image sizes to be applied to the DL

networks. The selection of 224x224 image size was made through a trial-and-error process in this

preprocessing stage, ensuring that reducing the size of CMR images does not lead to a reduction in

evaluation parameters. It was observed that reducing image sizes results in a decrease in hardware

resource usage. However, transformer models have a comparatively long training time when compared

to pretrained models. By reducing image size to 224x224, training time for these networks has

significantly decreased while simultaneously utilizing fewer hardware resources. In the continuation of

the pre-processing section, CutMix and MixUp techniques have been employed for data augmentation

(DA) [32-33]. This preprocessing approach represents the first novelty of this paper. In references [97-

100], review papers in the field of CVDs diagnosis from CMR images using AI techniques are reported.

It can be seen that this method has not been utilized for DA in similar research studies, and most

researchers have usually used GAN techniques for this task. In the following, the details of the CutMix

and MixUp methods are provided for the purpose of DA.

2.3. Data Augmentation

Data augmentation is a technique used in DL to increase the amount of training data by creating new

samples from existing ones [107-108]. This is done by applying a set of transformations to the original

data, creating variations that the model can learn from. DA is particularly useful when the amount of

training data is limited, as it can help prevent overfitting and improve model performance [51]. In

medical applications, researchers frequently encounter the challenge of limited access to medical

images containing a large number of subjects. To address this challenge, researchers have introduced

various methodologies, of which DA techniques are one of the most important [107-108]. References

[106-109] indicate that researchers have employed DA techniques such as different generative

adversarial network (GAN) models to diagnosis types of CVDs from CMR images. The results of these

studies demonstrate that the utilization of DA methods has led to an improvement in the accuracy and

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：

AutomaticDiagnosisofMyocarditisDiseaseinCardiacMRIModalityusingDeepTransformersandExplainableArtificialIntelligenceMahboobehJafari1,2,*,AfshinShoeibi1,2,NavidGhassemi1,JonathanHeras3,SaiHoLing4,AminBeheshti5,Yu-DongZhang6,Shui-HuaWang6,RoohallahAlizadehsani7,JuanM.Gorriz2,8,U.RajendraAcharya9,HamidA...

展开>> 收起<<

Automatic Diagnosis of Myocarditis Disease in Cardiac MRI Modality using Deep Transformers and Explainable Artificial Intelligence Mahboobeh Jafari12 Afshin Shoeibi12 Navid Ghassemi1 Jonathan Heras3 Sai Ho Ling4 Amin.pdf

共27页,预览5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！

Automatic Diagnosis of Myocarditis Disease in Cardiac MRI Modality using Deep Transformers and Explainable Artificial Intelligence Mahboobeh Jafari12 Afshin Shoeibi12 Navid Ghassemi1 Jonathan Heras3 Sai Ho Ling4 Amin

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: