Author: Dr. Raym Geis, member of the CAR AI Working Group
Artificial Intelligence (AI) and Machine Learning (ML) refer to computer algorithms that change their output based on exposure to more data input. These algorithms were initially described decades ago, but they became economically feasible to use only recently, because of more powerful computers combined with the availability of appropriate machine-consumable input data.
At their core, machine learning algorithms often are relatively simple, based on finding the maximum of a mathematical function. For building prototypes, ML computer code is compact, and much of it has already been written and is available in coding libraries, where one can download and combine the needed parts. It is surprisingly easy to build your own program to identify cats and dogs, and even identify your own dog among many others.
The power of these programs is their ability to find patterns in complex data. ML algorithms not only can find many of the same patterns as humans, but they detect sub-visual patterns, ones that humans just can’t see. Their economic power comes when applied to situations where they repeat the same procedure on multiple similar data inputs. Radiology examinations meet both requirements, of many similar cases, each with complex image data.
This ability to extract valuable new information from images, and do it more efficiently and reproducibly than now possible, will spark yet another major change in radiology practice, more than PACS or the advent of cross sectional imaging.
Despite the ease of building prototype ML algorithms, translating them into widely-used, clinically available products is difficult. Because of this, initial ML tools will be just that — tools. For image interpretation, these tools will fall into three classes of use cases:
- Separate normal from abnormal exams
- Improve computer aided lesion detection, described as “CAD that works,” often by decreasing false positives common in current CAD products.
- Radiomics, or feature extraction to improve disease description, predict lesion behavior, and drive precision medicine tailored to individual patients.
These tools will appear in two settings. First, ML tools will be built into OEM imaging machines. For example, for patients receiving an acute CT or MR for altered mental status, ML tools can screen for acute intracranial hemorrhage. While the scan is being performed, the tool can alert the radiologist that the scan is abnormal. Second, they will be options in reading workstations. Much as we now use 3D or multiplanar reconstruction tools, we can apply ML tools in appropriate settings to extract more information from images, such as textural differences in brain MRs of glioblastomas to predict the tumor’s genetic subtype.
In future posts, we will discuss how radiologists can develop, assess, and manage these tools.