Mitigating generalization errors of AI in image segmentation and classification for medical applications

Abstract

Deep learning-based systems in medical image analysis has encountered a significant limitation in the form of generalization errors in practical clinical applications. These errors, often due to insufficient labeled datasets covering the diverse spectrum of clinical conditions, limit the adaptability of these models across different clinical centers, vendors, and diseases. To address this challenge, my research aims to develop robust deep learning frameworks capable of mitigating overfitting and generalization issues. The primary focus of this work involves the development of generative model-based unsupervised learning framework for image segmentation. Moreover, the research delves into methodologies that enhance the generalizability of clinical decision support systems across diverse datasets. Additionally, the study explores the transfer of foundation model knowledge for more effective and generalized segmentation tasks and multi-modal clinical treatment plan delineation. This approach not only aims to address overfitting concerns but also enhance the precision and adaptability of deep learning models in medical imaging across varied clinical scenarios.