Teaching anatomy to a computer
Recent months have seen substantial advances in techniques for computer vision and image recognition. From Google demonstrations on YouTube videos to Baidu’s Deep Mind project, the machine learning technique known as “deep learning” has suddenly made it possible for computers to achieve image recognition accuracies far beyond prior capabilities. Indeed, Microsoft researchers in February published a paper reporting an error rate of just 4.94% on the popular ImageNet image set—this is below the estimated human error rate of 5.1%.
In this talk, I share how we and others are applying many of the same computational advances towards medical imaging, the challenges of doing so, and the rewards of success. I specifically will discuss substantial progress in the problem of anatomic segmentation, or precisely locating various anatomic structures, both gross and subtle, on CT scans, ultrasounds, and MR scans. This is a challenge that researchers have pursued for many years with mixed success. It is made difficult by the inherent complexity of fine 3D anatomy, the natural variability between patients, and differences in image quality.
Highly accurate anatomic segmentation has a variety of applications. I will walk through various examples--from clinical diagnosis to treatment planning to medical device development to research based on imaging big data. I will share my experience as a clinician spending countless hours contouring normal anatomy in radiation oncology treatment planning, and how new technology can spare physicians this burden. The technology can also potentially reduce medical oversights and spread access to best-practice medicine to underserved regions of the country and world.
I will also discuss how our team, bringing together pioneering computer scientists and physician researchers, is overcoming one of the key challenges of applying deep learning to a medical context, namely the need for large amounts of training data. I will discuss our efforts to build an open data set of tens of thousands of anonymized and annotated scans by recruiting a consortium of open-minded researchers and institutions. When complete, this imaging data set will be, to our knowledge, orders of magnitude larger than any previous open data set in medical imaging.
In many respects, the future imagined by computer vision researchers decades ago has finally arrived. The implications for medicine are great and, with computers finally learning anatomy, the next few years will be extremely exciting for researchers, physicians, and patients.