Andrew Hartz MD PGY2 and Michael Romano MS4
AI Education: A Deep Understanding of Deep Learning for Resident Radiologists
Artificial intelligence (AI) is currently a hot topic in radiology. AI refers to a large class of computer-based algorithms that perform functions similar to, or exceeding the limitations of, the human mind. Recently, a surge of Deep Learning applications has caught the attention of radiologists. Deep Learning algorithms, inspired by the function of the brain, consists of many layers of parameters that extract specific features from data in order to accomplish a task. Analogous to the many neurons that connect to form neural networks in the brain, Deep Learning consists of many parameters that connect to form artificial “neural networks” that are strengthened or weakened by “training” data. The combination of large amounts of digital data and the production of faster and more powerful computers (in particular, graphics processing units), has allowed for the rapid expansion of Deep Learning applications. Computer vision is one domain where Deep Learning has had extraordinary success. Deep Learning applications have been able to perform equal to or better than experienced radiologists in several image recognition tasks. Because image recognition is essential to radiology practice, it is important for radiologists and radiologists-in-training to understand how Deep Learning applications operate and how to incorporate them into improving radiology practice.
Our mission is to provide radiology residents with an educational platform to learn about AI and Deep Learning applications. In addition, we will provide a system where radiologists can use AI to supplement their own interpretations of radiographs and other modalities of imaging. Within this platform, there will be opportunities to test different applications and to contribute to the development of future applications.
A dedicated section on a Boston Medical Center / Boston University website will be used to host AI educational projects for radiologists. One of the first projects to be hosted includes an application inspired by Chester AI, the online diagnostic tool for chest radiographs. This application utilizes a publicly available Deep Learning model trained on multiple large chest radiograph datasets (including NIH Chest X-ray8 and RSNA Pneumonia Challenge), to create a probability table of diagnostic findings from uploaded chest radiographs by the user. Users will have the opportunity to test the application, explore the underlying code, and access additional educational resources.