Cited time in webofscience Cited time in scopus

Full metadata record

DC Field Value Language
dc.contributor.advisor 문인규 -
dc.contributor.author Jaewon Jeong -
dc.date.accessioned 2024-02-29T21:01:45Z -
dc.date.available 2024-03-01T06:00:31Z -
dc.date.issued 2024 -
dc.identifier.uri http://hdl.handle.net/20.500.11750/48074 -
dc.identifier.uri http://dgist.dcollection.net/common/orgView/200000730427 -
dc.description Deep learning cell morphology analysis;Deep learning cell morphology analysis;multi-class segmentation;marker control watershed;biomedical image analysis -
dc.description.abstract This paper proposes a deep learning-based method to automate muscle cell morphology segmentation and quantitative analysis in fluorescence images. The method employs U-Net and DeepLab-v3 architecture to effectively address challenges in automated myotube analysis. The goal of this study is to enhance the accuracy and efficiency of muscle cell and nuclei segmentation by refining the prediction of nuclei centers, contributing to the quantitative analysis of muscle cell growth characteristics.

To validate the proposed method, experiments were conducted using a dataset of fluorescence-stained images. The proposed method uses a modified deep learning model, specifically U-Net and DeepLab-v3, with additional layers for predicting nuclei centers. The model predicts myotubes and nuclei boundaries and centers, achieving more accurate segmentation of overlapping nuclei through marker-controlled watershed post-processing. Despite the challenges posed by the characteristics of myotube cell data, the proposed model's myotube cell segmentation yields superior results compared with conventional critical methods.

Furthermore, in response to issues arising limitations of U-Net structures from small foregrounds, such as class imbalances, we present a comparison with the results obtained by training with the DeepLab-v3 architecture. The DeepLab-v3 model exhibits exceptional performance in precisely segmenting nuclei, particularly addressing the challenges faced by U-Net.

Therefore, this study introduces a novel myotube morphology analysis method that integrates deep learning and marker-controlled watershed algorithms. The proposed method, validated through experiments and comparisons, demonstrates the potential to overcome the limitations of in existing segmentation techniques.|본 논문은 형광 이미지에서 근육세포 형태 분할 및 정량적 분석을 자동화하기 위해 딥 러닝 기반 방법을 제안한다. 제안된 방법은 U-Net 및 DeepLab-v3 아키텍처를 활용하여 자동화된 근육세포 분석과 관련된 문제를 효과적으로 해결한다. 이는 근육세포와 핵 분할의 정확성과 효율성을 향상하고, 핵의 중앙 마커를 개선하여 근육세포 성장 특성의 정량적 분석에 기여하는 것을 목표로 한다.

제안한 방법의 검증을 위해 형광 염색된 이미지 데이터셋을 이용하여 실험을 진행했다. 제안된 방법은 딥 러닝, 특히 핵 중심 예측을 위한 추가 레이어가 있는 U-Net과 DeepLab-v3 기반으로 수정된 모델을 활용한다. 모델은 근육세포와 핵 경계 및 핵 중심을 예측한 후 마커 제어 유역 사후 처리를 통해 중첩되는 핵을 보다 정확하게 분할한다. 근육세포 데이터 특성으로 인한 어려움에도 제안된 모델의 근육세포 분할은 기존 임계 방식 보다 뛰어난 결과를 보여준다.

또한, 작은 전경으로 인한 문제와 클래스 불균형 같은 U-Net 구조의 한계로 인해, DeepLab-v3-v3 아키텍처로 훈련한 결과를 함께 비교한다. DeepLab-v3 모델은 특히 U-Net이 어려움을 겪는 문제에서 핵을 정확하게 분할하는 데 탁월한 성능을 보여준다.

따라서, 본 연구에서는 딥 러닝과 마커 제어 유역 알고리즘을 결합한 새로운 근육세포 형태분석 방법을 제시한다. 실험과 비교를 통해 검증된 제안 방법은 기존 분할 기법의 한계를 극복하는 데 가능성을 보여준다.
-
dc.description.tableofcontents Ⅰ. Introduction
1.1 Research Background & Contribution 1
1.2 Related Work 2
Ⅱ. Data Generation 5
Ⅲ. Deep Learning Methods
3.1 U-Net 7
3.2 Selection of DeepLab-v3 Over U-Net Structure Limitations 9
3.3 DeepLab-v3 10
3.4 Marker-controlled Watershed Algorithm 11
IV. Model Objective 13
V. Result
5.1 Myotube Segmentation 16
5.2 Nuclei Segmentation 18
5.3 Evaluation Metrics 21
5.3.1 IoU(Intersection over Union) 21
5.3.2 Dice Coefficient 21
5.3.3 Precision 21
5.3.4 Recall 22
5.3.5 Mean IoU(mIoU), Mean Dice 22
5.3.6 Cell Count Accuracy (%) 22
5.4 Quantitative Result 23
Ⅵ. Conclusion 26
-
dc.format.extent 30 -
dc.language eng -
dc.publisher DGIST -
dc.title Automated Multi-Class Segmentation for Myotube Morphology in Fluorescent Images Using Deep Learning -
dc.type Thesis -
dc.identifier.doi 10.22677/THESIS.200000730427 -
dc.description.degree Master -
dc.contributor.department Department of Robotics and Mechatronics Engineering -
dc.contributor.coadvisor Youhyun Kim -
dc.date.awarded 2024-02-01 -
dc.publisher.location Daegu -
dc.description.database dCollection -
dc.citation XT.RM정73 202402 -
dc.date.accepted 2024-01-30 -
dc.contributor.alternativeDepartment 로봇및기계전자공학과 -
dc.subject.keyword Deep learning cell morphology analysis -
dc.subject.keyword multi-class segmentation -
dc.subject.keyword marker control watershed -
dc.subject.keyword biomedical image analysis -
dc.contributor.affiliatedAuthor Jaewon Jeong -
dc.contributor.affiliatedAuthor Inkyu Moon -
dc.contributor.affiliatedAuthor Youhyun Kim -
dc.contributor.alternativeName 정재원 -
dc.contributor.alternativeName Inkyu Moon -
dc.contributor.alternativeName 김유현 -
Files in This Item:

There are no files associated with this item.

Appears in Collections:
Department of Robotics and Mechatronics Engineering Theses Master

qrcode

  • twitter
  • facebook
  • mendeley

Items in Repository are protected by copyright, with all rights reserved, unless otherwise indicated.

BROWSE