نبذة مختصرة : sažetak (hrvatski): Ovaj se rad bavi rekonstrukcijom slika magnetske rezonance mozga primjenom metoda dubokog učenja implementiranih u fastMRI biblioteku. Magnetska rezonanca je ključna metoda za di- jagnostiku u medicini. Cilj ovog rada je istražti primjene tehnike prijenosa učenja na unaprijed istreniranim U-Net i End-to-End VarNet modelima kako bi se poboljšala kvaliteta rekonstrukcije slika uz poduzorkovane podatke. Korišteni su fastMRI i Calgary-Campinas skupovi podataka, a podatci su pripremljeni uz primjenu različitih maski poduzorkovanja ekvidistantnih, slučajnih ekvidistantnih i maski s Gaussovim šumom. Modeli su trenirani i evaluirani pomoću SSIM i PSNR metrika. Rezultati pokazuju da prijenosom učenja možemo poboljšati kvalitetu rekons- trukcije, pogotovo na End-to-End VarNet modelu. Kvaliteta rekonstrukcija još uvijek ne doseže razinu potrebnu za kliničku upotrebu. ; This paper addresses the reconstruction of magnetic resonance images using deep learning met- hods implemented in fastMRI library. Magnetic resonance is a key diagnostic tool in medicine. The aim of this paper is to research uses of transfer learning techniques on pre-trained U-Net and End-to- End VarNet models to improve the quality of image reconstruction from undersam- pled data. fastMRI and Calgary-Campinas datasets were used, and data was prepared using multiple undersampling masks: equispaced, random equispaced and masks with Gaussian no- ise. The models were trained and evaluated using SSIM and PSNR metrics. Results show that transfer learning can enhance reconstruction quality, particularly for the End-to-End VarNet model. However, the quality of reconstructions still falls short of the level required for clinical use.
Processing Request
No Comments.