BROWSE

ITEM VIEW & DOWNLOAD

Multi-Scale Approaches for Pathology Image Analysis

DC Field Value Language
dc.contributor.advisor Chun, Se Young -
dc.contributor.author Hong, WonJae -
dc.date.accessioned 2020-04-01T08:56:40Z -
dc.date.available 2020-04-01T08:56:40Z -
dc.date.issued 2020-02 -
dc.identifier.other 200000288723 -
dc.identifier.uri http://unist.dcollection.net/common/orgView/200000288723 en_US
dc.identifier.uri https://scholarworks.unist.ac.kr/handle/201301/31788 -
dc.description Department of Electrical Engineering -
dc.description.abstract The multi-scale approach has been applied to deep learning methods for image processing such as image segmentationand image super-resolution. So we researched the impact of using the multi-scale information in terms of image segmentation and image generation in this thesis. In lesion segmentation, Segmentation of skin lesions is often done using image segmentation as an important pre-processing step before skin melanoma images are analyzed. With the advent of U-Net in the field of medical image segmentation, it is possible to segment skin lesions of superior quality. We would like to propose a structure that can use multi-scale information more effectively based on U-Net. Specifically, we devised a deep neural network that combines various U-Nets operating in multi-scale to obtain final segmentation results through U-Nets that perform final skin lesion segmentation. We have examined the advantages and disadvantages of the proposed multi-scale method by combining experiments at various resolutions. In image genetation research, We made various attempts to generate pathological images using SinGAN with a multiscale pyramid pipeline structure. During training, SinGAN are learning to generate down-sampled input image in the initial steps, so it is focused on global features. And, as scale goes up, it is focused on fine areas. We applied them to various image classifications using GAN-generated images. Through patient classification experiments, we attempted to identify the effect of preventing the leakage of personal information of concern from pathological images and identified about 30% degradation in classification performance. And through the Tumor classification experiment, we checked the possibility of using the GAN- generated image as data to train the model, and confirmed its value as training data by having the same performance as real pathology images. However, since the two experiments are conducted on a small sample, additional experiments are required after collecting more data. -
dc.description.statementofresponsibility close -
dc.language ENG -
dc.publisher Graduate School of UNIST -
dc.title Multi-Scale Approaches for Pathology Image Analysis -
dc.type Master's thesis -
dc.administration.regnum 200000288723 -
Appears in Collections:
EE_Theses_Master

find_unist can give you direct access to the published full text of this article. (UNISTARs only)

Show simple item record

qrcode

  • mendeley

    citeulike

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

MENU