Date of Award
Doctor of Science in Information Systems
Deep learning has driven AI's rapid growth in recent years, especially in the medical domain, where deep CNNs are the state-of-the-art for image recognition and classification. However, training them from scratch is challenging due to the lack of data and high computational requirements. Transfer Learning (TL) is an effective approach for limited training data, and TL integrated with GANs has improved image analysis models. It was unclear how much impact big data-driven Deep Learning systems had on adoption and acceptance in real-world healthcare. Specifically, the effectiveness of recently developed DL systems as scalable and generalizable AI applications remained an open question.
Accordingly, the main objective of this research work is to assess the effectiveness of TL-GAN systems on broad adoption. This study explored the combination of transfer learning and generative adversarial networks (GANs) in medical imaging by conducting a systematic literature review. In addition, the scalability dimension of these systems was evaluated by examining the dynamics of GAN-augmented datasets and the accuracy achieved on target datasets. Finally, the generalization capabilities of the combination of transfer learning and GANs were evaluated.
The study added to the current literature on TL and GANs in medical imaging, specifically in image synthesis and computational efficiency. Two strategies for combining TL and GANs were identified and summarized. The study also examined the impact of artificially augmented training datasets on the Fine-Tuning layer, finding that larger datasets resulted in more parameters being trained for optimal performance. Additionally, the study investigated the effect of synthetic dataset size on classification accuracy in TL settings, concluding that target validation accuracy stabilized as the dataset size increased. Furthermore, the study explored the generalizability of the models trained on GAN-augmented datasets and found that pre-trained models exhibited good performance when applied to various target datasets, indicating a high degree of generalizability in the models.
Sutrave, Kruttika, "TOWARDS LONG-TERM IMPACT OF DEEP LEARNING SYSTEMS IN MEDICAL IMAGING" (2023). Masters Theses & Doctoral Dissertations. 428.