Difficulty Measure for Radiology Cases with use of Item Analysis and Deep Learning

A vast and extensive radiology education is fundamental for the diagnosis, monitoring, and treatment of lung cancer. Image synthesis, with Generative Adversarial Networks (GANs), can be a powerful tool in radiology education by its ability to diversify training cases for medical students and radiology residents. By controlling the image synthesis, images can be produced with a specific difficulty or complexity level that fits the student’s level. To achieve optimal personalization of education, the knowledge gap should be defined by a concrete measure. This research focuses on a measure of difficulty for the detection of lung nodules. Item analysis is a statistical method that can give an indication of the difficulty based on the responses of a group of individuals. To automate the calculation of a measure of difficulty, deep neural networks are used to perform item analysis on lung nodule cases. The method is validated by comparing the measure of difficulty with a subjective subtlety score given by experienced radiologists. The ordinal logistic regression analysis shows a statistically significant relationship between the calculated measure of difficulty and the subtlety scores of nodules. A measure of difficulty is defined that has the potential to be applied to image synthesis for the design of computer-assisted learning systems.