Segmentacija fibroze srca s pomočjo konvolucijskih avtokodirnikov

ŠTUHEC, TIM (author), Žabkar, Jure (mentor)

V nalogi obravnavamo problem segmentacije fibroze srca. Na voljo imamo 200 simulacij MRF slik srca, med katerimi so samo 3 slike srca brez fibroze. Posnemamo realno stanje, kjer je pridobivanje segmentacij slik srca zamudno in je večina src, slikanih z magnetno resonanco, bolnih. Problema smo se lotili s pomočjo avtokodirnikov, za najboljše so se izkazali konvolucijski. Konvolucijske nevronske mreže smo uporabili na dva načina. V prvem smo s slikami poskusili rekonstruirati identične slike brez fibroze, v drugem pa smo poskusili lokalizirati samo fibrozo. Drugi način se je izkazal kot veliko uspešnejši, saj dosega dobre rezultate, medtem pa je imel prvi težave zaradi premajhnega števila slik zdravega srca. Kljub temu prva metoda odpira več možnosti za nadaljnje raziskovanje na tem področju, saj ne potrebuje slik s priloženimi segmentacijami, ampak le podatke o tem, katere slike predstavljajo zdravo srce in katere srce s fibrozo.

• Language: Slovenian
• Keywords: fibroza srca, konvolucijske nevronske mreže, avtokodirniki, samonadzorovano učenje
• Work type: Bachelor thesis/paper
• Typology: 2.11 - Undergraduate Thesis
• Organization: FRI - Faculty of Computer and Information Science ; FMF - Faculty of Mathematics and Physics
• Year: 2022
• PID: 20.500.12556/RUL-140834
• COBISS.SI-ID: 124604675
• Publication date in RUL: 19.09.2022

Secondary language

• Language: English
• Title: Segmentation of cardiac fibrosis with convolutional autoencoders

Abstract

In the assignment, we deal with the problem of heart fibrosis segmentation. We have 200 MRF simulations of cardiac images available, among which only 3 cardiac images without fibrosis. We simulate a real world situation where obtaining segmentations of heart images is time-consuming and most hearts imaged with magnetic resonance are diseased. We tackled the problem with the help of autoencoders, convolutional ones turned out to be the best. We used convolutional neural networks in two ways. In the first, we tried to reconstruct identical images without fibrosis, and in the second, we tried to localize only the fibrosis. The second method turned out to be much more successful, achieving good results, while the first had its problems due to the insufficient number of images of a healthy heart. Nevertheless, the first method offers more opportunities for further research in this area, since it does not require images with attached segmentations, but only the information about, which images represent a healthy heart and which represent a heart with fibrosis.

• Keywords: cardiac fibrosis, convolutional neural networks, autoencoders, self-supervised learning