Segmentation of Multiple Sclerosis Lesions across Hospitals Learn Continually or Train from Scratch Enamundram Naga Karthik12Anne Kerbrat3Pierre Labauge4

2025-05-03 0 0 3.39MB 7 页 10玖币

侵权投诉

Segmentation of Multiple Sclerosis Lesions across

Hospitals: Learn Continually or Train from Scratch?

Enamundram Naga Karthik1,2∗Anne Kerbrat3Pierre Labauge4

Tobias Granberg5Jason Talbott6Daniel S. Reich7Massimo Filippi8,9

Rohit Bakshi10 Virginie Callot11,12 Sarath Chandar 2,13 Julien Cohen-Adad1,2,14

1NeuroPoly Lab, Polytechnique Montreal, Canada 2MILA, Quebec AI Institute

3Neurology Department, CHU Rennes, Rennes, France

4MS Unit, Department of Neurology, CHU Montpellier, Montpellier, France

5Department of Clinical Neuroscience, Karolinska Institutet, Stockholm, Sweden

6Department of Radiology and Biomedical Imaging, Zuckerberg San Francisco

General Hospital, University of California, San Francisco, CA, USA

7National Institute of Neurological Disorders and Stroke, National Institutes of Health, USA

8Neuroimaging Research Unit, Institute of Experimental Neurology, Division of

Neuroscience, and Neurology Unit, IRCCS San Raffaele Scientiﬁc Institute, Milan, Italy

9Vita-Salute San Raffaele University, Milan, Italy

10Brigham and Women’s Hospital, Harvard Medical School, Boston, USA

11AP-HM, CHU Timone, Pole de Neurosciences Cliniques, Department of Neurology,

Marseille, France 12Aix-Marseille Univ, CNRS, Marseille, France

13Department of Computer and Software Engineering, Polytechnique Montreal, Canada

14Functional Neuroimaging Unit, CRIUGM, Université de Montréal, Montréal, Canada

Abstract

Segmentation of Multiple Sclerosis (MS) lesions is a challenging problem. Sev-

eral deep-learning-based methods have been proposed in recent years. However,

most methods tend to be static, that is, a single model trained on a large, spe-

cialized dataset, which does not generalize well. Instead, the model should learn

across datasets arriving sequentially from different hospitals by building upon

the characteristics of lesions in a continual manner. In this regard, we explore

experience replay, a well-known continual learning method, in the context of MS

lesion segmentation across multi-contrast data from 8 different hospitals. Our

experiments show that replay is able to achieve positive backward transfer and

reduce catastrophic forgetting compared to sequential ﬁne-tuning. Furthermore,

replay outperforms the multi-domain training, thereby emerging as a promising

solution for the segmentation of MS lesions. The code is available at this link.

1 Introduction

Multiple Sclerosis (MS) is a chronic, neurodegenerative disease of the central nervous system. Lesion

segmentation from magnetic resonance images (MRI) serves as an important biomarker in measuring

disease activity in MS patients. However, manual segmentation of MS lesions is a tedious process,

hence motivating the need for automated tools for segmentation. Several deep-learning (DL) based

methods have been proposed in the past few years [

]. They tend to be trained in a static manner - all

the datasets are pooled, jointly preprocessed, shufﬂed (to ensure they are independent and identically

distributed, IID) and then fed to the DL models. While this has its beneﬁts, it does not represent

∗Corresponding Author: naga-karthik.enamundram@mila.quebec

36th Conference on Neural Information Processing Systems (NeurIPS 2022).

arXiv:2210.15091v1 [cs.CV] 27 Oct 2022

Figure 1: Overview of our meth-

ods. Four experiments were per-

formed - A:Single-domain train-

ing: a model is trained individu-

ally on each center. B:Sequen-

tial ﬁne-tuning: after training the

model on center n, the pre-trained

encoder weights are loaded for

center n+1 (red dashed arrows).

C:Experience replay: in addition

to ﬁne-tuning (as in B) upto 20

samples per each center are stored

in the memory buffer (in gray). D:

Multi-domain training: data from

all centers are pooled and a joint

model is trained.

a realistic scenario. First, it is difﬁcult to pool datasets from multiple hospitals with increasing

privacy concerns. Second, since MS is a chronic disease, one would imagine a scenario where a DL

model, like humans, engages in continual learning (CL) [

] and builds upon the lesion characteristics

from different centers when presented sequentially. However, this sequential knowledge acquisition

presents a major problem in DL models known as catastrophic forgetting [4].

Previous works in CL for medical imaging have used regularization-based [

] and memory-

based methods [

] for tackling catastrophic forgetting. In this work, we formalize the MS lesion

segmentation across multiple hospitals as a domain-incremental learning problem [

], where the

task remains unique (i.e. segmentation of lesions) but the model is sequentially presented with new

domains (i.e. data from different hospitals). Four types of experiments are performed as shown in

Figure 1 - single-domain,multi-domain sequential ﬁne-tuning, and experience replay. Our results

show that replay helps in reducing catastrophic forgetting and achieves positive backward transfer,

that is, the segmentation performance on data seen earlier improves as the model continues to learn

sequentially. Furthermore, we show that replay outperforms multi-domain training as more data

arrive sequentially, thereby suggesting that the CL is a better long-term solution than re-training the

model from scratch on a large, curated dataset.

2 Experience Replay for MS Lesion Segmentation

Replay (or, rehearsal) presents a straightforward way to prevent catastrophic forgetting and improve

the performance on new domains. Let xdenote the patches of the 3D volumes, ybe the corresponding

labels,

fθ

denote the neural network with parameters

, and

denote the joint dataset. In a standard

IID training regime, the loss Lis given by Equation 1.

L=E(x,y)∼D[`(fθ(x), y)]

(1) L0=E(x,y)∼Dk[`(fθ(x), y)] + E(x,y)∼M [`(fθ(x), y)] (2)

In this work, we use the simplest form of experience replay wherein training data from previously

encountered domains

are stored in a memory buffer and interleaved with the current domain’s

training data. Particularly, the dataset

is divided into 8 different domains (

D1,D2, . . . D8

). The

model is trained sequentially on one dataset

at a time. For each dataset

(

1≤k < 8

), we store

upto 20 image-label pairs (depending on the dataset size) in a memory buffer

and merge them

with the training data of the current domain. The updated loss term

is given by Equation 2. Due to

unconstrained access to the multi-center data, the model is tested on all the remaining centers once it

has been trained on one center. We use the Dice Loss [9] as the loss function `.

SoftSeg

To account for the partial volume effects at the lesion boundaries, we use soft ground

truth labels in our training procedure. In addition to mitigating the partial volume effects, soft

segmentations [

] were shown to generalize better and to reduce model-overconﬁdence. In this

We use domains and centers interchangeably. A domain is essentially a center (i.e. a hospital) that

holds/provides the data. Hence, data from each new center is treated as a different 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：

SegmentationofMultipleSclerosisLesionsacrossHospitals:LearnContinuallyorTrainfromScratch?EnamundramNagaKarthik1;2AnneKerbrat3PierreLabauge4TobiasGranberg5JasonTalbott6DanielS.Reich7MassimoFilippi8;9RohitBakshi10VirginieCallot11;12SarathChandar2;13JulienCohen-Adad1;2;141NeuroPolyLab,PolytechniqueMon...

展开>> 收起<<

Segmentation of Multiple Sclerosis Lesions across Hospitals Learn Continually or Train from Scratch Enamundram Naga Karthik12Anne Kerbrat3Pierre Labauge4.pdf

共7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！

Segmentation of Multiple Sclerosis Lesions across Hospitals Learn Continually or Train from Scratch Enamundram Naga Karthik12Anne Kerbrat3Pierre Labauge4

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: