|
Abstract:
|
In recent years, deep learning models have demonstrated strong potentials in medical images analysis. However, different from natural image analysis where huge amount training samples can be easily collected and analysis is typically on the "coarse" level, the medical images that are available for training are usually limited and usually we need to focus on fine-grained very detailed level information. This makes direct deployment of models that are trained on natural images not suitable. In this talk, I will summarize the recent efforts on developing high-performance machine learning models for medical image analysis including: Effective integration with domain knowledge; Knowledge transfer across the models trained on multiple related domains; Federated learning to enable multi-institution collaborations with privacy preservation; Integration with other complementary but relevant information such as genetics and clinical records; Robust interpretation of the black-box models. I will explain the motivation and principle behind those research, with representative examples. Finally I will envision how the future machine learning pipeline for medical image analysis will look like.
|
