Deep Learning Mixture -of-Experts Approach for Cytotoxic Edema Assessment in Infants and Children

2025-05-06 0 0 953.19KB 14 页 10玖币

侵权投诉

Deep Learning Mixture-of-Experts Approach for

Cytotoxic Edema Assessment

in Infants and Children

Henok Ghebrechristos1∗, Stence Nicholas MD 2∗,

David Mirsky MD 2∗, Gita Alaghband PhD 1, Manh Huynh1, Zackary Kromer1, Ligia Batista 2

, Brent O’Neill MD2, Steven Moulton MD2, Daniel M. Lindberg MD 2

University of Colorado Denver, Department of Computer Science and Engineering1

University of Colorado Denver, School of Medicine2

Abstract

This paper presents a deep learning framework for image classification aimed at increasing predictive

performance for Cytotoxic Edema (CE) diagnosis in infants and children. The proposed framework includes

two 3D network architectures optimized to learn from two types of clinical MRI data – a trace Diffusion

Weighted Image (DWI) and the calculated Apparent Diffusion Coefficient map (ADC). This work proposes

a robust and novel solution based on volumetric analysis of 3D images (using pixels from time slices) and

3D convolutional neural network (CNN) models. While simple in architecture, the proposed framework

shows significant quantitative results on the domain problem. We use a dataset curated from a Children’s

Hospital Colorado (CHCO) patient registry to report a predictive performance F1 score of 0.91 at

distinguishing CE patients from children with severe neurologic injury without CE. In addition, we perform

analysis of our system’s output to determine the association of CE with Abusive Head Trauma (AHT) – a

type of traumatic brain injury (TBI) associated with abuse – and overall functional outcome and in-hospital

mortality of infants and young children. We used two clinical variables, AHT diagnosis and Functional

Status Scale (FSS) score, to arrive at the conclusion that CE is highly correlated with overall outcome and

that further study is needed to determine whether CE is a biomarker of AHT. With that, this paper introduces

a simple yet powerful deep learning-based solution for automated CE classification. This solution also

enables an in-depth analysis of progression of CE and its correlation to AHT and overall neurologic

outcome, which in turn has the potential to empower experts to diagnose and mitigate AHT during early

stages of a child’s life.

1 INTRODUCTION

1.1 TRAUMATIC BRAIN INJURY

Traumatic Brain Injury (TBI) in children causes more than 2,000 deaths, 35,000 hospitalizations, and

470,000 emergency department visits in the US each year, making it a leading cause of pediatric disability

and death [2]. The youngest children also face the additional risk of Abusive Head Trauma (AHT), a

particularly deadly type of TBI that is easy to miss because caregivers rarely provide an accurate history,

victims are typically pre-verbal infants, and clinical examination findings are subtle or non-specific [3]–

[5].

AHT is the most important source of morbidity and mortality for abused children. The rate of AHT has

been shown consistently to be 20-36/100,000 per year in the first year of life, and to be far and away the

largest source of abusive mortality. Outcomes for children with AHT are dismal, with mortality rate of

approximately 20% and permanent neurological sequelae in approximately 80% of survivors [6] – [9].

Victims of AHT frequently require ongoing medical support including speech, physical, and occupational

therapy [5]. In the first 4 years alone, the medical costs for each case of AHT average roughly $50,000 [6].

The mechanisms and pathophysiology that differentiate AHT from other forms of TBI are poorly

understood. While repetitive, rotational forces (as occur in shaking) are thought to be key to why AHT has

worse outcomes compared to non-inflicted TBI (niTBI), this remains controversial, and the

pathophysiological processes are not well understood [1]. Secondary injury – when damage continues after

the initial trauma ends – is hypothesized to be more important in AHT than other forms of TBI, but it too

is poorly understood [4], [20]. Cytotoxic Edema (CE) – a pathological process occurring when energy-

requiring cell membrane ion pumps fail resulting in an abnormal influx of water into the cell leading to cell

death – is thought to be central to secondary injury [20], [22] – [24].

CE has been well-described in cases of ischemic stroke and hypoxic brain injuries in adults, children, and

neonates [25], [26]. AHT has been shown to produce patterns of CE similar to those seen in hypoxic-

ischemic injuries, suggesting that hypoxia (either during the abusive episode itself or as a result of trauma-

related apnea) is an important mediator of secondary injury [27]. CE is also thought to be more common in

AHT than in other forms of TBI and to be a marker of poor prognosis [23], [28] – [30]. If CE proves to be

a reliable biomarker of AHT, the problem of misdiagnosing AHT as niTBI could be mitigated in early

stages. CE, if it is confirmed as specific to AHT, could also enhance understanding of the pathophysiology

at work in AHT and how it differs from other forms of trauma.

No prior studies have assessed in a robust, systematic way the characteristics of CE in pediatric TBI. The

proposed solution, which uses deep learning techniques to extract CE patterns form data by incorporating

radiology workflow, could supply an evidence base to understand the significance of CE in pediatric TBI.

More importantly, if CE provides an early predictor of neurologic outcome, this information could affect

AHT treatment. Currently, the treatments of AHT are largely supportive care and monitoring for seizures

[7], [8]. Better understanding of the pathophysiology of AHT, particularly the secondary injury component,

which commonly occurs once a patient is under medical care where therapeutic interventions could be

administered, could lead to new treatments.

We postulate that CE can help discriminate AHT from niTBI

and help predict the neurologic outcome – two tasks essential

to recognizing AHT and protecting the child. To that end, the

development of an automated assistive tool or algorithm to

classify MR-based imaging data, such as structural MRI

data, and, more importantly, to distinguish subjects with

features of CE from healthy subjects, can aid in not only

understanding patterns of CE appearance in MRI data but

also to diagnose and mitigate abuse-related head trauma in

early stages. In this paper, we present a robust deep learning

framework and algorithms that considers radiology

workflow. The main contributions of this work can be listed

as follows:

Day 1

Day 3

Figure 1. As cells swell due to inward shift of water,

there is a commensurate decrease in diffusion,

identified as high signal on DWI (right) and low signal

on ADC (left).

• We present a novel deep learning framework and network architectures optimized to learn patters

of Cytotoxic Edema from volumetric time slices of brain MRI of infants and children.

• We evaluate the performance of the proposed solution and present results that demonstrate the

efficacy of the framework in predicting the presence or absence of Cytotoxic Edema using clinical

data of infants and children less than 10 years of age admitted to Children’s Hospital Colorado

(CHCO) [9].

1.2 CYTOTOXIC EDEMA IMAGE CHARACTERISTICS

Cytotoxic edema (CE), a type of cerebral edema, is the result of cells being unable to maintain ATP-

dependent sodium/potassium (Na+/K+) membrane pumps which are responsible for high extracellular and

low intracellular Na+ concentration [10]. When an insult to the brain results in ischemia or hypoxia, this

oxygen-requiring process produces very little to no new ATP. Cells quickly use up their reserves of ATP

and, unless normoxia (normal oxygen level) is restored, the cellular machinery loses its ability to sustain

homeostasis. This results in intracellular swelling and reduction in the extracellular volume, the

combination of which causes a reduction in the observed degree of isotropic water diffusion. The trace DWI

is generated by summing the signal of water diffusion in 3 perpendicular directions in space. This trace

DWI image displays areas of abnormally reduced water diffusion as hyperintense (bright), but it is also T2

weighted and thus can be confounded by areas that are also T2 hyperintense. Therefore, an ADC map is

usually also automatically created by the scanner. This is a parametric map of the signal intensity of a purely

T2* weighted image subtracted from the trace DWI. Voxels in this ADC map therefore represent the

calculated, actual ADC values at a given voxel, which by convention are become darker with greater

degrees of diffusion restriction. Thus, areas of restricted diffusion indicating CE are represented on imaging

as bright or hyperintense on DWI, and dark or hypointense on ADC (Figure 1). For details on clinical and

imaging characteristics of CE, refer to publications by Liang et. al. [11] and Rosenblum [10].

The rest of the paper is organized as follows:

• We first present related work and relevant research outcomes on the domain problem;

• We then present the method of study where we detail study population, dataset preparation and

system design;

• We then follow that section by sharing our results and discussion before we present conclusion

remarks and future direction of this research.

2 RELATED WORKS

Advances in Magnetic Resonance Imaging (MRI) enabled the non-invasive visualization of the infant’s

brain through acquired high-resolution images [12], [13]. The increasing availability of large-scale datasets

of detailed infant brain multi-modal MR images (e.g., T1-weighted (T1w), T2-weighted (T2w), diffusion-

weighted MRI (dwMRI), and resting-state functional MRI (rsfMRI) images) affords unique opportunities

to accurately study early Abusive Head Trauma (AHT), leading to insights into the origins and abnormal

developmental trajectories of CE.

Changes in brain structure and function caused by CE and their correlation to TBI, morbidity, and overall

outcome have proved of great interest to research groups. In diagnostic imaging of adults in particular,

machine learning based classification and predictive modeling of the stages of Cerebral Edema [14] in large

cohorts of TBI patients have been investigated [15]– [18].

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：

DeepLearningMixture-of-ExpertsApproachforCytotoxicEdemaAssessmentinInfantsandChildrenHenokGhebrechristos1∗,StenceNicholasMD2∗,DavidMirskyMD2∗,GitaAlaghbandPhD1,ManhHuynh1,ZackaryKromer1,LigiaBatista2,BrentO’NeillMD2,StevenMoultonMD2,DanielM.LindbergMD2UniversityofColoradoDenver,DepartmentofComputerS...

展开>> 收起<<

Deep Learning Mixture -of-Experts Approach for Cytotoxic Edema Assessment in Infants and Children.pdf

共14页,预览3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！

Deep Learning Mixture -of-Experts Approach for Cytotoxic Edema Assessment in Infants and Children

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: