Unmet Need
Advances in medical imaging has radically changed the practice of medicine in the last few decades. Recently, machine learning techniques have been shown to produce higher fidelity images in various modalities and to aid in diagnostics (Lundervold & Lundervold, 2019). Machine learning imaging research require large datasets to be trained in order to operate effectively. Presently, datasets for human studies depend on digital phantoms such as the Zubal and XCAT phantoms. However, these are based on average human anatomical data and lack the ability to produce variations in the human body that can be seen in clinics. These phantoms can be supplemented with real anatomical imaging data such as those from Magnetic Resonance Imaging (MRI), but machine learning models often depend on thousands of images – an infeasible number to obtain in the clinic. Therefore, there is a strong need for a method to generate large and sufficiently representative datasets of medical images for use in training machine learning medical imaging applications.
Technology Overview
Johns Hopkins researchers have developed an artificial intelligence method for generating a realistic digital human phantom population that reflects the anatomical variability of both normal and abnormal patient populations with potentially unlimited numbers. Utilizing multiple algorithms that generate synthetic examples and attempt to discriminate real from synthetic examples, researchers are able to rapidly produce large datasets of medical images for use in other applications.
Stage of Development
Currently in the process of refining the machine learning algorithm.
Publication
Wenyi Shao, Kevin Leung, Steven Rowe, Martin Pomper, Yong Du. Digital brain phantoms by generative adversarial network (GAN). Journal of Nuclear Medicine May 2021, 62 (supplement 1) 1545; https://jnm.snmjournals.org/content/62/supplement_1/1545