Skip to main content

LUP Student Papers

LUND UNIVERSITY LIBRARIES

Deep-learning based semantic segmentation of healthy and metaplastic airway epithelium in lung tissue samples

Brunnmark, Sven LU and Sjögren, Filip LU (2021) In Master's Thesis in Mathematical Sciences FMSM01 20211
Mathematical Statistics
Abstract (Swedish)
I detta examensarbete utforskas möjligheten att använda så kallade "fully convolutional neural networks" för att utföra semantisk segmentering av epitelvävnad och metaplastisk epitelvävnad i hematoxylin-infärgade lungvävnadsprover. Vidare söker avhandlingen avgöra vilken nätverks-arkitektur som bäst löser detta problem. De arkitekturer som undersöks är ett flertal variationer av nätverket U-net samt ett antal DeepLabV3+, vilka modifierats med olika "encoders", kodare, närmare bestämt MobileNetV2, Xception, ResNet-18 och ResNet-50. Datan som används i projektet består av ultra-högupplösta bilder av lungvävnad innehållande cirka 1-1.5 miljarder pixlar. För epitelmodelleringen finns 32 sådana bilder att tillgå och för den metaplastiska... (More)
I detta examensarbete utforskas möjligheten att använda så kallade "fully convolutional neural networks" för att utföra semantisk segmentering av epitelvävnad och metaplastisk epitelvävnad i hematoxylin-infärgade lungvävnadsprover. Vidare söker avhandlingen avgöra vilken nätverks-arkitektur som bäst löser detta problem. De arkitekturer som undersöks är ett flertal variationer av nätverket U-net samt ett antal DeepLabV3+, vilka modifierats med olika "encoders", kodare, närmare bestämt MobileNetV2, Xception, ResNet-18 och ResNet-50. Datan som används i projektet består av ultra-högupplösta bilder av lungvävnad innehållande cirka 1-1.5 miljarder pixlar. För epitelmodelleringen finns 32 sådana bilder att tillgå och för den metaplastiska epitelmodellering finns 13, där de sistnämnda härstammar från patienter som avlidit i COVID-19. Utöver detta finns för varje bild tillhörande delinjeringar av den intressanta vävnaden. Till epitelbilderna är dessa delinjeringar handritade men för COVID-19-bilderna består de av immunhistokemiska markörer. Resultaten av modelleringen visade att DeepLabV3+-modellen med MobileNetV2-kodare var mest effektiv och uppnådde ett genomsnittligt IoU på 0.9207, precision på 0.8903 och recall på 0.9411 vid applicering på testdata-setet. För den metaplastiska epitelmodelleringen visade tester att ingen av modellerna lyckades uppnå acceptabla resultat, vilket kan härledas till brister i de kemiskt infärgade delinjeringsmaskerna samt för stora variationer i färgbalans och kontrast mot epitelbilderna för att de tidigare modellerna skulle ha möjlighet att överföras på COVID-19-datan. (Less)
Abstract
In this thesis, we explore the possibility of using fully-convolutional neural networks to semantically segment epithelium and metaplastic epithelium in hematoxylin-stained lung tissue samples. In addition, the thesis aims to find which network architecture is most effective for this task. The networks tested included several variations of the U-net network and various architectures converted to a fully-convolutional network using the DeeplabV3+ structure, including MobileNet V2, Xception, ResNet-18, and ResNet-50. The data used included ultra-high resolution images of lung tissue samples containing roughly 1-1.5 billion pixels each, with 32 images for epithelium modelling, and 13 images from COVID-19 patients for metaplastic epithelium... (More)
In this thesis, we explore the possibility of using fully-convolutional neural networks to semantically segment epithelium and metaplastic epithelium in hematoxylin-stained lung tissue samples. In addition, the thesis aims to find which network architecture is most effective for this task. The networks tested included several variations of the U-net network and various architectures converted to a fully-convolutional network using the DeeplabV3+ structure, including MobileNet V2, Xception, ResNet-18, and ResNet-50. The data used included ultra-high resolution images of lung tissue samples containing roughly 1-1.5 billion pixels each, with 32 images for epithelium modelling, and 13 images from COVID-19 patients for metaplastic epithelium modelling. In addition to this, labels for training and evaluation the performance for the models were also provided, with manually-delineated masks for the epithelium data and immunohistochemically-stained masks for the metaplastic epithelium data.
It was shown that the DeepLabV3+ architecture using MobileNetV2 was the most effective for epithelium modelling, achieving a mean IoU of $0.9207$, a precision of $0.8903$, and a recall of $0.9411$ on the testing data-set. For metaplastic epithelium modelling however, the model was not able to semantically segment the epithelium in an acceptable or practically useful manner. This was mainly due to the limitations of the chemically-stained masks, as they did not include consistent or particularly relevant features for the classification of metaplastic epithelium. Furthermore, variations in the color balance and contrast of the images compared to the epithelium data rendered transfer learning of the previous models ineffective. (Less)
Please use this url to cite or link to this publication:
author
Brunnmark, Sven LU and Sjögren, Filip LU
supervisor
organization
course
FMSM01 20211
year
type
H2 - Master's Degree (Two Years)
subject
publication/series
Master's Thesis in Mathematical Sciences
report number
LUTFMS-3415-2021
ISSN
1404-6342
other publication id
2021:E21
language
English
id
9044534
date added to LUP
2021-07-02 11:43:23
date last changed
2022-02-02 16:22:28
@misc{9044534,
  abstract     = {{In this thesis, we explore the possibility of using fully-convolutional neural networks to semantically segment epithelium and metaplastic epithelium in hematoxylin-stained lung tissue samples. In addition, the thesis aims to find which network architecture is most effective for this task. The networks tested included several variations of the U-net network and various architectures converted to a fully-convolutional network using the DeeplabV3+ structure, including MobileNet V2, Xception, ResNet-18, and ResNet-50. The data used included ultra-high resolution images of lung tissue samples containing roughly 1-1.5 billion pixels each, with 32 images for epithelium modelling, and 13 images from COVID-19 patients for metaplastic epithelium modelling. In addition to this, labels for training and evaluation the performance for the models were also provided, with manually-delineated masks for the epithelium data and immunohistochemically-stained masks for the metaplastic epithelium data.
 It was shown that the DeepLabV3+ architecture using MobileNetV2 was the most effective for epithelium modelling, achieving a mean IoU of $0.9207$, a precision of $0.8903$, and a recall of $0.9411$ on the testing data-set. For metaplastic epithelium modelling however, the model was not able to semantically segment the epithelium in an acceptable or practically useful manner. This was mainly due to the limitations of the chemically-stained masks, as they did not include consistent or particularly relevant features for the classification of metaplastic epithelium. Furthermore, variations in the color balance and contrast of the images compared to the epithelium data rendered transfer learning of the previous models ineffective.}},
  author       = {{Brunnmark, Sven and Sjögren, Filip}},
  issn         = {{1404-6342}},
  language     = {{eng}},
  note         = {{Student Paper}},
  series       = {{Master's Thesis in Mathematical Sciences}},
  title        = {{Deep-learning based semantic segmentation of healthy and metaplastic airway epithelium in lung tissue samples}},
  year         = {{2021}},
}