Weak supervision with imperfect annotations for medical image computing

点击翻译按钮获取中文摘要

Medical image computing (MIC) aims to develop computational methods for solving problems pertinent to medical image interpretation and understanding. Recent advances in machine learning have made a big leap in MIC and we have started to witness their influence on our daily lives, ranging from 3D human embryo/organ visualization to intelligent examination of lung scans for COVID-19 checkup. Moreover, the healthcare industry is increasingly embracing the cloud services powered by machine learning techniques, such as Google Cloud healthcare and IBM Watson Health. In the future, the envisioned MIC solutions are expected to free clinicians from tedious daily operations on easy cases, to provide physicians a second opinion, and to facilitate biomedical researchers for numerical biomarker discovery. Meanwhile, there are still many hurdles to overcome. In most real-world MIC tasks, it is extremely difficult to attain high quality datasets that can be used in strong supervision methods due to the high cost of medical data labeling; in contrast, weak annotation, or imperfect annotation, is much easier to access. The imperfect annotation in various MIC tasks makes machine learning techniques working with weak supervision to be highly desired. With my long-term goal to leverage machine learning techniques and computer vision algorithms to improve medical imaging relevant healthcare, this research proposal aims to innovate different weak supervision algorithms to extract medical information that is not explicitly available in imperfect annotations. Three different topics will be the major directions in this proposal, including: (1) learning anomaly from sparse, extremely unbalanced annotated medical images, through innovating representation learning methods that fuse batch contrastive learning and generative learning; (2) learning from noisy labeled medical images, by incorporating Bayesian inference into deep learning towards a unified framework; (3) inference from imprecision annotated medical images, by developing a unified weak-supervision region-of-interest segmentation pipeline that combines class activation map methods and multiple instance learning. Potential impact of this proposal includes (1) inventing solutions for medical image computing with imperfect annotation, which will open a door for the massive existing medical images to become directly usable in MIC research and (2) proposing representation learning methods that will be helpful to discover numerical biomarkers of abnormal, rare cases. Though our target application domain is medical image computing, the solutions proposed in this research are also transferable to many image analysis and computer vision tasks. Besides, the HQP trained through this research program will get training on both scientific theories and application-oriented developments, that would be valuable in their future medical imaging and machine learning careers.