Detekcija metastatskog tkiva na patohistološkim slikama presjeka limfnih čvorova uporabom dubokog učenja

master's thesis

• Title (english): Metastatic tissue detection in histopathological images of lymph node section using deep learning
• Author: Andrija Miličević
• Mentor: Gorana Aralica (mentor)
• Committee member: Sven Seiwerth (predsjednik povjerenstva)
• Committee member: Slavko Gašparov (član povjerenstva)
• Committee member: Gorana Aralica (član povjerenstva)
• Granter: University of Zagreb; School of Medicine; (Department of Pathology); Zagreb
• Defense date and country: 2019-07-12, Croatia
• Scientific / art field, discipline and subdiscipline: BIOMEDICINE AND HEALTHCARE; Clinical Medical Sciences; Pathology

Abstract

Algoritmi umjetne inteligencije omogućuju računalima obavljanje inteligentnih zadataka.
Veliki doprinos umjetne inteligencije očekuje se u medicini gdje bi ovakvi algoritmi
mogli postati standard u svakodnevnom radu liječnika. Patologija je grana medicine koja
se bavi proučavanjem bolesti, a velik dio posla odnosi se na mikroskopsku dijagnostiku bolesti,
što znači da patolozi rade sa slikama. Duboke neuronske mreže inspirirane su načinom
kako mozak obrađuje informacije, a posebno su prikladne za rad sa slikama. U ovom radu
predložena je metoda za detekciju metastatskog tkiva na patohistološkim slikama presjeka
limfnih čvorova, a temelji se na dubokim neuronskim mrežama. Predložena metoda postiže
točnost od 86,57% na PatchCamelyon skupu podataka, a uspoređena je s drugim objavljenim
metodama. PatchCamelyon je skup podataka od 327 680 slika u boji, veličine 96 x 96. Za
svaku sliku, označeno je postojanje metastatskog tkiva u središnjem dijelu slike. Ovaj skup
podataka nastao je nasumičnim izrezivanjem segmenata 400 preparata iz Camelyon16 skupa
podataka. Augmentacija PatchCamelyon skupa podatka značajno utječe na točnost predložene
metode. Korištenje većeg skupa podataka za učenje moglo bi dodatno povećati točnost
predložene metode. Ovom metodom moguće je obraditi 250 slika u sekundi na prosječnom
računalu. Podizanjem praga klasifikacije moguće je povećati preciznost predložene metode i
koristiti je za brzu detekciju metastaza nad cijelim slikama patohistoloških presjeka limfnih
čvorova. Ova metoda mogla bi ubrzati rad patologa označavanjem regija preparata koji nisu
dio normalnog tkiva limfnog čvora.

Abstract (english)

Artificial intelligence algorithms allow computers to perform intelligent tasks. Big impact
is expected in medicine where such algorithms could become a standard in the day-today
work of doctors. Pathology is a branch of medicine that studies disease. A big part of
it is related to microscopic disease diagnostics, which means pathologist work with images.
Deep neural networks are inspired by the way the brain processes information and are particularly
suitable for working with images. In this thesis, a method for detecting metastatic
tissue in histopatological images of lymph node sections using deep neural networks is proposed.
The proposed method achieves 86,57% accuracy on the PatchCamelyon data set and
is compared with other published methods. PatchCamelyon is a collection of 327 680 color
images with dimensions 96 x 96. For each image, there are annotations of metastatic tissue
presence in the central image region. This data set was generated by extracting segments
of 400 whole-slide images from the Camelyon16 data set. Augmentation of the PatchCamelyon
data set significantly influences the accuracy of the proposed method. Using a larger
training data set could increase the accuracy of this method. This method can be used on
an average computer in real-time, as it can process 250 images per second. By increasing
the classification threshold, it is possible to increase the accuracy of the proposed method so
it can be used for fast detection of metastases in whole-slide images of lymph nodes. This
method could accelerate the work of pathologists by marking regions of a whole-slide image
which are not part of the normal lymph node tissue.

• Keywords: detekcija metastatskog tkiva; duboke neuronske mreže; patologija
• Keywords (english): metastatic tissue detection; deep neural networks; patology
• Language: croatian
• URN:NBN: urn:nbn:hr:105:790669
• Study programme: Title: Medicine; Study programme type: university; Study level: integrated undergraduate and graduate; Academic / professional title: doktor/doktorica medicine (doktor/doktorica medicine)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Open access
• Created on: 2020-02-11 08:51:33