Rekonstrukcija oblika organa iz para rendgenskih slika korištenjem neuronskih mreža

master's thesis

• Title (english): Reconstruction of organ shapes from pairs of X-ray images using neural networks
• Author: Dominik Sekula
• Mentor: Matko Orsag (mentor)
• Committee member: Matko Orsag (član povjerenstva)
• Committee member: Tamara Petrović (član povjerenstva)
• Committee member: Frano Petric (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2024-02-27, Croatia

Abstract

Ovaj rad istražuje primjenu Neural Radiance Fields (NeRF) za problem rekonstrukcije trodimenzionalnih modela iz dvodimenzionalnih rendgenskih slika. Korištenjem generiranih digitalno rekonstruiranih radiografa iz AbdomenCT-1K skupa podataka, analizirane su različite metode rekonstrukcije radi evaluacije performansi na rekonstrukciji modela. MedNeRF metoda, posebno prilagođena medicinskoj rekonstrukciji korištena je za potrebe izvedbe ovog rada. Evaluacija modela izvodila se tijekom treniranja putem FID i KID mjera kvalitete modela, dok pri rekonstrukciji koristile su se mjere PSNR i SSIM. Rezultate predstavljamo kao analizu optimalne količine slika po pacijentu, ali i usporedbu između standardne MedNeRF implementacije koristeći i 1 sliku za rekonstrukciju, naspram one gdje koristimo 2 slike i novu funkciju gubitka.

Abstract (english)

This paper investigates the application of Neural Radiance Fields (NeRF) to the problem of reconstructing three-dimensional models from two-dimensional X-ray images. Using generated digitally reconstructed radiographs from the AbdomenCT-1K dataset, different reconstruction methods were analyzed to evaluate the model’s performance. The MedNeRF method, specially adapted to medical reconstruction, was used to conduct this work. Model evaluation was performed during training using FID and KID quality measures, while PSNR and SSIM measures were used during reconstruction. We present the results as an analysis of the optimal amount of images per patient, but also a comparison between the standard MedNeRF implementation using 1 image for reconstruction, versus the one proposed in the thesis, where we use 2 images and a new loss function.

• Keywords: NeRF; 3D rekonstrukcija; 3D model; polje zračenja; DRR
• Keywords (english): NeRF; medicine; 3D reconstruction; 3D model; radiance fields; DRR
• Language: croatian
• URN:NBN: urn:nbn:hr:168:368863
• Study programme: Title: Information and Communication Technology; Study programme type: university; Study level: graduate; Academic / professional title: magistar/magistra inženjer/inženjerka informacijske i komunikacijske tehnologije (magistar/magistra inženjer/inženjerka informacijske i komunikacijske tehnologije)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Open access
• Created on: 2024-03-28 07:54:15