Availability of training datasets
To function properly, machine learning algorithms in radiology need to be trained on large amounts of medical images. The more, the better. But in the medical field, it is difficult to gain access to such datasets. For the sake of comparison, a typical non-medical imaging dataset can contain up to 100,000,000 images, while medical imaging sets rarely exceed about 1,000 images.
Another problem is producing labeled datasets for supervised training. Medical image annotation is a very time-consuming and labor-intensive process. Radiologists and other medical experts must do this task manually assigning appropriate labels for the given AI application. There is potential for automatically extracting structured labels from radiology reports using natural language processing. But even then, radiologists will most likely need to review the results.
Opting for existing algorithms instead of developing custom ones can also be problematic. Many successful deep learning models available on the market are trained on 2D images, while CT scans and MRIs are 3D. This extra dimension poses a problem, and the algorithms need to be adjusted.
Finally, AI technology itself is leaving room for doubt. Computer power has been doubling every two years. However, according to Wim Naude, a business professor from the Netherlands, this established pattern is diminishing. Consequently, we may not have the necessary power and multitasking abilities to take over the broad range of tasks that an average radiologist performs. AI’s silicon-based transistors will have to be replaced with technology such as organic biochips, which is still in its infantry to achieve such capabilities.