Classifying non-small cell lung carcinoma in histological images using a convolutional neural network
Timonen, Veera (2019)
2019
Bioteknologian tutkinto-ohjelma
Lääketieteen ja terveysteknologian tiedekunta - Faculty of Medicine and Health Technology
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2019-05-24
Julkaisun pysyvä osoite on
http://urn.fi/URN:NBN:fi:tuni-201907152578
http://urn.fi/URN:NBN:fi:tuni-201907152578
Tiivistelmä
The purpose of this Master’s thesis was to teach a convolutional neural network to recognize non-small cell lung cancer from whole slide images (WSI) and to separate regions of interest from other tissue. IHC fluoro-chromogenically stained whole slide images under brightfield illumination were used as target images, and the same WSIs with cytokeratin masks applied under fluorescent illumination were used as input images.
An immunohistochemical fluoro-chromogenic dye is done when PD-L1-expressing tumor regions and PD1-expressing alveolar macrophages need to be distinguished. Cytokeratin-positive carcinoma regions show clearly in brightfield images. It is important to separate these regions especially when considering immunotherapy as treatment, because there exist antibody based medications against both PD1- and PD-L1 expressing tumor- and lymphocyte cells, and the areas surrounding cancer may cause false positives leading to immunotherapy being poorly targeted.
The method is based on U-net architecture in a convolutional neural network. A CNN is capable of achieving excellent results in tasks including image recognition, and U-net has been specifically designed for medical image analysis tasks.
The results show that the neural network used is capable of distinguishing cancer regions from other tissue with good accuracy (AUC = 0.96).
An immunohistochemical fluoro-chromogenic dye is done when PD-L1-expressing tumor regions and PD1-expressing alveolar macrophages need to be distinguished. Cytokeratin-positive carcinoma regions show clearly in brightfield images. It is important to separate these regions especially when considering immunotherapy as treatment, because there exist antibody based medications against both PD1- and PD-L1 expressing tumor- and lymphocyte cells, and the areas surrounding cancer may cause false positives leading to immunotherapy being poorly targeted.
The method is based on U-net architecture in a convolutional neural network. A CNN is capable of achieving excellent results in tasks including image recognition, and U-net has been specifically designed for medical image analysis tasks.
The results show that the neural network used is capable of distinguishing cancer regions from other tissue with good accuracy (AUC = 0.96).