Image Prior based Adversarial Self-supervised Feature Learning and its Applications in Semantic Understanding
Mar 09,2022 Projects
Project description/goals
Develop self-supervised learning algorithms and pretrained models for specific visual data domain.
Importance/impact, challenges/pain points
Medical image datasets lack enough annotated samples for fully-supervised model training.
Solution description
Take advantages of adversarial learning and prior knowledge of specific medical image data to design self-supervised learning algorithms, and reduce the annotation cost for model training. For examples, we discard some image patch of the input and train an encoder-decoder model to reconstruct the patches, which pretrains the encoder in a self-supervised learning manner.
Key contribution/commercial implication
Brain metastases detection, Brain MRI analysis, Histopathological image analysis.
Next steps
Develop a multi-modal self-supervised pre-training algorithm for brain metastases detection.
Develop a self-supervised adversarial learning method for semi-supervised brain MRI segmentation.
