This is Not a Real Image: Generative Artificial Intelligence to Enhance Radiology Education

Radiologists fulfill a critical role in our healthcare system, but their workload has increased substantially over time. Although algorithmic tools have been proposed to support the diagnostic process, the workload is not efficiently decreased in this manner. However, another possibility is to decrease workload in a different area. The main topic of this thesis is concerned with investigating how simulation training can be realized to aid in the image interpretation skills training of the radiologist and radiology resident. To realize simulated training it is necessary to know (1) how we can create realistic artificial medical images, subsequently (2) how we can control their variety and (3) how we can adjust their diagnostic difficulty.

Generative artificial intelligence techniques are used to create the artificial medical images. Initially, a Generative Adversarial Network type architecture is used to create 2-dimensional artificial medical images. The model can guide coarse features, i.e. larger anatomical structures, of the images. The created artificial images are assessed both quantitatively and qualitatively in terms of their realism and results show they can blend in with original medical images. A diffusion model type architecture is used to better control the finer features, i.e. smaller anatomical structures like pathology, of the artificial medical images. The results show that the model was able to adjust fine-feature characteristics of the pathology type, in addition to the coarse-feature guidance.

Furthermore, a method is presented to describe the diagnostic difficulty of an (artificial) medical image. It is shown that it is possible to create new artificial images with pathology and image characteristics that pertain to a certain diagnostic difficulty level. Additionally, the responsible implementation of a ‘medical image simulator’ to assist in image interpretation skills is investigated. To this end, two extreme scenarios are posed to identify possible harm; one where image interpretation skills are trained solely using traditional training methods, and one where practicing with cases is fully replaced by the medical image simulator.

Combining the results of this thesis resulted in a prototype of such a 'medical image simulator'. This simulator can take over part of the workload of the supervising radiologists, by providing a means for independent repetitive practice for the resident. Furthermore, it can also decrease the workload of the resident. The realistic artificial medical images can be varied in terms of their content and their difficulty. This can enable a personalized experience that can enhance training of image interpretation skills and make it more efficient.