Large Batch and Patch Size Training for Medical Image Segmentation Junya Sato and Shoji Kido

2025-05-03 2 0 4.03MB 9 页 10玖币

侵权投诉

Large Batch and Patch Size Training

for Medical Image Segmentation?

Junya Sato and Shoji Kido

Osaka University Graduate School of Medicine, 2-2 Yamadaoka, Suita-city, Osaka

565-0871, Japan

{j-sato,kido}@radiol.med.osaka-u.ac.jp

Abstract. Multi-organ segmentation enables organ evaluation, accounts

the relationship between multiple organs, and facilitates accurate diag-

nosis and treatment decisions. However, only few models can perform

segmentation accurately because of the lack of datasets and computa-

tional resources. On AMOS2022 challenge, which is a large-scale, clinical,

and diverse abdominal multiorgan segmentation benchmark, we trained

a 3D-UNet model with large batch and patch sizes using multi-GPU

distributed training. Segmentation performance tended to increase for

models with large batch and patch sizes compared with the baseline set-

tings. The accuracy was further improved by using ensemble models that

were trained with diﬀerent settings. These results provide a reference for

parameter selection in organ segmentation.

Keywords: Medical image analysis ·Organ segmentation ·Distributed

learning.

1 Introduction

Automatic organ segmentation is critical in analyzing human anatomy. In addi-

tion to a qualitative assessment by the radiologist, quantitative assessment using

automatic segmentation assists in the accurate diagnosis of diseases. Multi-organ

segmentation can assess organ interactions and provide more detailed informa-

tion than single-organ segmentation. Although multiorgan segmentation plays

a vital role in computer-aided diagnosis, it suﬀers from computational-memory

limitations. Moreover, because CT and MRI images acquired in daily clinical

practice comprise high-resolution 3D images, this limitation makes it even more

diﬃcult to input whole images into deep learning models. Current methods usu-

ally resize images to smaller sizes at the expense of organ details or crop a portion

of the image (patch) and input it into the model without using the positional

information of the surrounding organs. Some studies have shown that a large

patch size allows a model to make more accurate predictions [2,4].

?This work was partly achieved through the use of SQUID at the Cybermedia Center,

Osaka University.

This study was supported by JSPS KAKENHI (grant numbers: JP21H03840).

arXiv:2210.13364v1 [eess.IV] 24 Oct 2022

2 J. Sato and S. Kido.

The batch size is also a critical parameter that inﬂuences training eﬀective-

ness. Larger batch sizes are associated with more accurate gradient estimates

[6]; however, batch size is also aﬀected by memory limitations. In this study,

the model could only be trained with a smaller batch size for the 3D organ

segmentation compared to tasks, such as natural image recognition.

Therefore, we hypothesize that increasing the patch and batch sizes would

improve the training accuracy. To test this hypothesis, we used an nnU-Net [6]

model, which is the gold standard for medical image segmentation. The dataset

used was the Abdominal Multi Organ Segmentation 2022 (AMOS2022) challenge

[7], which consists of 500 CT and 100 MRI scans collected from multiple sites,

a wide range of imaging conditions, and patient backgrounds with 15 voxel-

level annotations. We trained the model with distributed learning using multiple

GPUs and searched for the batch and patch sizes with the best segmentation

accuracy.

2 Method

2.1 Approach

In this study, the preprocessing, network architecture, training strategy, and

postprocessing were based on the default nnU-Net conﬁguration. As referred to in

the original paper, a larger batch size enables more accurate gradient estimates,

and a larger patch size increases contextual information among organs [6]. Loss

function is another factor that signiﬁcantly aﬀects the accuracy of the training

process. We compared changes in accuracy by varying these parameters.

2.1.1 Training with Large Patch Size

nnU-Net initializes the patch size to the median image shape and iteratively

reduces it while adapting the network topology accordingly, until the network can

be trained with a batch size of at least 2. For example, on an NVIDIA RTX3090,

the patch size was automatically set to (depth,height,width) = (80,160,160) for

both Task1 and Task2 by the default nnU-Net conﬁguration. We set the patch

size to [64,160,160], [64,288,288], [64, 354,354], and [64,384,384]. In all cases, the

batch size was set to 2.

2.1.2 Training with Large Batch Size

Multi-organ segmentation using a 3D deep learning model is often performed

with a batch size of 1-4[1,15,16,9,13], which is smaller than the batch size used

for learning typical 2D images (usually 16 or more)[3,5]. The default nnU-Net

conﬁguration requires a batch size of at least 2. In our experiments, we varied

the batch size from 2 to 16.

2.1.3 Loss Function

Compound loss allows for robust predictions in medical image segmentation[12].

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：

LargeBatchandPatchSizeTrainingforMedicalImageSegmentation?JunyaSatoandShojiKidoOsakaUniversityGraduateSchoolofMedicine,2-2Yamadaoka,Suita-city,Osaka565-0871,Japanfj-sato,kidog@radiol.med.osaka-u.ac.jpAbstract.Multi-organsegmentationenablesorganevaluation,accountstherelationshipbetweenmultipleorgans,...

展开>> 收起<<

Large Batch and Patch Size Training for Medical Image Segmentation Junya Sato and Shoji Kido.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！

Large Batch and Patch Size Training for Medical Image Segmentation Junya Sato and Shoji Kido

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: