Skip to main content

LUP Student Papers

LUND UNIVERSITY LIBRARIES

Deep Learning with a DAG Structure for Segmentation and Classification of Prostate Cancer

Flood, Gabrielle LU (2016) In Master's Theses in Mathematical Sciences FMA820 20161
Mathematics (Faculty of Engineering)
Abstract
Deep learning is a machine learning technique inspired by the biological nervous system. The method has been used more and more over the last decades. Within this thesis, convolutional neural networks (CNN:s) are used for Gleason classification of prostate cancer in histopathological images. The data that has been used consists of digitalized microscopic images of prostate biopsies stained with haematoxylin-eosin (H&E). There are separate samples containing benign, Gleason 3, Gleason 4 and Gleason 5 tissue. More specifically, the potential of a directed acyclic graph (DAG) structure for the classifying CNN has been investigated. In this network a segmentation image has been used as a second label in the middle of the network, and the... (More)
Deep learning is a machine learning technique inspired by the biological nervous system. The method has been used more and more over the last decades. Within this thesis, convolutional neural networks (CNN:s) are used for Gleason classification of prostate cancer in histopathological images. The data that has been used consists of digitalized microscopic images of prostate biopsies stained with haematoxylin-eosin (H&E). There are separate samples containing benign, Gleason 3, Gleason 4 and Gleason 5 tissue. More specifically, the potential of a directed acyclic graph (DAG) structure for the classifying CNN has been investigated. In this network a segmentation image has been used as a second label in the middle of the network, and the backpropagation has thus been performed using two different outputs. Hence, as a buildup, a CNN with a simple structure was created for segmentation of different components in benign prostate tissue. The structure and trained weights of this segmentation network were then used in the classifying DAG net. No final fine-tuning of the hyper-parameters or optimization of the network has been performed, but the cost function showed a decreasing trend when the training was carried out. Therefore it was concluded that the DAG structure does have potential and that further development of the method could yield a high performance network. (Less)
Please use this url to cite or link to this publication:
author
Flood, Gabrielle LU
supervisor
organization
course
FMA820 20161
year
type
H2 - Master's Degree (Two Years)
subject
keywords
Deep learning, convolutional neural networks, CNN, directed acyclic graph, DAG, prostate cancer
publication/series
Master's Theses in Mathematical Sciences
report number
LUTFMA-3304-2016
ISSN
1404-6342
other publication id
2016:E44
language
English
id
8890727
date added to LUP
2016-09-12 13:52:02
date last changed
2016-09-12 13:52:02
@misc{8890727,
  abstract     = {{Deep learning is a machine learning technique inspired by the biological nervous system. The method has been used more and more over the last decades. Within this thesis, convolutional neural networks (CNN:s) are used for Gleason classification of prostate cancer in histopathological images. The data that has been used consists of digitalized microscopic images of prostate biopsies stained with haematoxylin-eosin (H&E). There are separate samples containing benign, Gleason 3, Gleason 4 and Gleason 5 tissue. More specifically, the potential of a directed acyclic graph (DAG) structure for the classifying CNN has been investigated. In this network a segmentation image has been used as a second label in the middle of the network, and the backpropagation has thus been performed using two different outputs. Hence, as a buildup, a CNN with a simple structure was created for segmentation of different components in benign prostate tissue. The structure and trained weights of this segmentation network were then used in the classifying DAG net. No final fine-tuning of the hyper-parameters or optimization of the network has been performed, but the cost function showed a decreasing trend when the training was carried out. Therefore it was concluded that the DAG structure does have potential and that further development of the method could yield a high performance network.}},
  author       = {{Flood, Gabrielle}},
  issn         = {{1404-6342}},
  language     = {{eng}},
  note         = {{Student Paper}},
  series       = {{Master's Theses in Mathematical Sciences}},
  title        = {{Deep Learning with a DAG Structure for Segmentation and Classification of Prostate Cancer}},
  year         = {{2016}},
}