1. RSUIR at Rangsit University
2. RSU Digital Collection
3. วิทยาลัยวิศวกรรมชีวการแพทย์ COLLEGE OF BIOMEDICAL ENGINEERING
4. BioEng-BE-D-Thesis

Title:	Synthetic data for deep learning medical applications : generation, evaluation, and utilization
Authors:	Asadi, Fawad
metadata.dc.contributor.advisor:	Jamie A. O'Reilly
Keywords:	Generative Adversarial Networks;Performance -- Evaluation;Computational intelligence;Artificial intelligenc
Issue Date:	2024
Publisher:	Rangsit University
metadata.dc.description.other-abstract:	Deep learning algorithms show promise in medicine for tasks like diagnosis, treatment planning, and health monitoring due to their ability to analyze diverse data and detect complex patterns, though their development is hindered by constraints such as the need for large, labeled, high-quality, and unbiased training datasets, posing challenges in their collection and preparation. Generative adversarial networks (GANs) can potentially address challenges. This research involved training progressively growing GAN (PGGAN) to generate synthetic computed tomography (CT) images of the lungs, investigating the impact of weight initialization methods on Inception v3 performance, and employing StyleGAN2 with adaptive discriminator augmentation (ADA) to generate synthetic fluid-attenuated inversion recovery (FLAIR) magnetic resonance (MRI) images with corresponding masks for deep learning training. The initial findings indicated that synthetic images generated by the PGGAN resembled real images, but there was room for improvement in GAN's performance. The subsequent study emphasized the superiority of pre-trained weights over randomly initialized models. Lastly, StyleGAN2-ADA performed well with a Fréchet inception distance (FID) score of 14.39. The addition of synthetic images did not significantly impact Unets' performance, but geometric augmentation alongside synthetic images enhanced generalization. Overall, despite their modest impact on training, synthetic data holds the potential to address data collection challenges, warranting broader exploration across other deep learning applications beyond segmentation tasks.
Description:	Dissertation (D.Eng. (Biomedical Engineering)) -- Rangsit University, 2023
metadata.dc.description.degree-name:	Doctor of Engineering
metadata.dc.description.degree-level:	Doctoral Degree
metadata.dc.contributor.degree-discipline:	Biomedical Engineering
URI:	https://rsuir-library.rsu.ac.th/handle/123456789/2833
metadata.dc.type:	Thesis
Appears in Collections:	BioEng-BE-D-Thesis

Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.