Federated Learning and Meta-Learning for Medical Image Classification: Effective Aggregation Techniques and Adaptive Learning Rate Approaches
Year: 2024
Type: dissertation
Abstract: Computer-aided diagnosis (CAD) has the potential to revolutionize medical diagnosis through advanced\nmachine learning (ML) algorithms. However, the efficacy of CAD systems relies heavily on extensive, diverse,\nand annotated datasets for robust training. This dissertation proposes a multifaceted strategy to address data\nscarcity, privacy concerns, and diverse data distribution across healthcare institutions by harnessing two major\ntypes of learning methods: federated learning (FL) and model-agnostic meta-learning (MAML). In the first\nlearning method, FL enables CAD systems to train on diverse data across multiple institutions, overcoming\nindividual dataset limitations. However, data heterogeneity (e.g., due to non-independent and identically\ndistributed data) among clients causes client model divergence, making it challenging to achieve convergence\nduring model aggregation. To address these challenges, we propose different aggregation techniques: randomsampling federated maximum (FedRSMax), random-sampling federated median (FedRSMed), and federated\nrank-based client selection aggregation (FedRSA). In the second learning method, MAML has the potential to\nfacilitate swift adaptation to varying medical imaging datasets, effectively navigating discrepancies in imaging\nprotocols and disease presentations. However, MAML struggles with fixed learning rates, limiting adaptation\nto varying task complexities. To address this challenge, we propose a task similarity-driven meta-adaptive\nlearning rate mechanism. Overall, this research pursues a two-fold strategy: (1) to develop an efficient\naggregation technique for medical image classification using FL; and (2) to employ MAML within a few-shot\nlearning framework, optimizing adaptation to datasets with limited samples. The proposed methodologies\nare evaluated using a comprehensive set of performance metrics, including accuracy, convergence rate, and\nadaptation loss. By integrating these cutting-edge techniques, this dissertation advances the frontier of CAD\nsystems, paving the way for more accurate, adaptable, and scalable solutions in medical image analysis.
Author: Hossen, Nazmul
Institution: University of Saskatchewan
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W6991290907
