Predtreniranje dubokih neuronskih mreža na analizu medicinskih slika

master's thesis

• Title (english): Pretraining of deep neural networks used for medical image analysis
• Author: Petar Cvitanović
• Mentor: Sven Lončarić (mentor)
• Committee member: Sven Lončarić (predsjednik povjerenstva)
• Committee member: Marko Subašić (član povjerenstva)
• Committee member: Tomislav Petković (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2024-07-15, Croatia

Abstract

Nakon transplantacije srca, moguća su odbacivanja transplantanta. Proces odbaciva-
nja presađenog organa bitno je detektirati i tretirati ako je potrebno. Za tu svrhu, biop-
sijom i histološkom analizom prikupljene su slike srca nakon transplantacije na kojima
je vidljiva struktura srčanog tkiva. Kroz rad je implementiran model neuronske mreže
treniran na skupu slika srca kako bi mogao prepoznati odbacivanje i klasificirati ga u
jednu od četiri klase. Model neuronske mreže također je proširen multiple instance lear-
ning metodom koja dozvoljava da model treniramo na skupu slika, umjesto pojedinač-
nim slikama, za što imamo pripremljen skup podataka. Modeli s i bez multiple instance
learning metode su nakon treniranja evaluirani i uspoređeni po svojim F1 mjerama i
vrijednostima funkcije gubitka.

Abstract (english)

After a heart transplant, graft rejections are possible. The process of rejection of the
transplanted organ is essential to detect and treat if necessary. For this purpose, biopsy
and histological analysis were used to collect images of the heart after transplantation,
showing the structure of the heart tissue. The paper explains the implementation of
a neural network model trained to recognize rejection and classify it into one of four
classes depending on the rejection severity. The neural network model is also extended
with the multiple instance learning method, which allows us to train the model on a set
of images instead of individual images, for which we have a prepared data set. Models
with and without the multiple instance learning method were evaluated after training
and compared according to their F1 measures and loss function values.

• Keywords: strojno učenje; računalni vid; neuronska mreža; multiple instance learning; transplantacija srca 31
• Keywords (english): machine learning; computer vision; neural network; multiple instance learning; heart transplant
• Language: croatian
• URN:NBN: urn:nbn:hr:168:217654
• Study programme: Title: Computing; Study programme type: university; Study level: graduate; Academic / professional title: magistar/magistra inženjer/inženjerka računarstva (magistar/magistra inženjer/inženjerka računarstva)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Closed access
• Created on: 2024-10-12 20:18:46