Radiologists are trained to diagnose diseases from medical images. One of these diseases is haemorrhage, which is usually assessed via a Computed Tomography (CT) scan. Before having enough experience to do this, they will follow a trajectory as a radiologist trainee to build up diagnostic skills. However, radiologist trainees are mostly trained on cases that are currently being assessed/treated or cases that the supervisor has available. This results in a varying degree of experiences of these graduated radiologists. Also, in their future career, their diagnostic skills are dependent on the cases they encounter. To level out their diagnostic experience, they would benefit from a systematic approach to practice with a variety of cases.
Seventy deep neural networks (DNNs) are trained on a database of ~3000 CT-slices to classify them into either a haemorrhage or a non-haemorrhage case. These DNNs vary in their classifying performance. Some cases may be classified correctly by all the networks, whereas some will only be classified correctly by one network. The goal of the assignment is to exploit the varying performance of these networks, to research the possibility to use their outcomes to systematically order the cases. Using this information it may be possible to create a small version of a Computer Assisted Learning (CAL) system. Such a system should systematically showcase to the user/trainee in order to give users/trainees the same skill level and experience at the end of the training. It might also be necessary to take into account the difficulty of a case, to start with easier images and gradually move towards more difficult ones.