Radiologists are trained to diagnose diseases from medical images. One of these diseases is lung cancer, which is usually assessed using a Computed Tomography (CT) scan. During their training, they gain experience by practicing on cases that are currently being treated or on cases that the supervisor has available. This results in a varying degree of experiences of graduated radiologists. To level out their diagnostic experience, they would benefit from a systematic method to practice with a variety of cases.
To achieve such a systematic ordering of cases, it might be possible to use multiple Deep Learning (DL) networks. Suppose we have several DL networks, all trained for the diagnosis of lung cancer, with varying degrees of performance. It is hypothesised that we can exploit the varying performance of these networks to systematically order cases. If all networks have a correct diagnosis, the case could be labelled an ‘easy’ case. Conversely, if none of the networks have a correct diagnosis, the case could be labelled a ‘difficult’ case.
The information on the “difficulty” of a case gives an opportunity to create a Computer Assisted Learning system (CAL). Such a system should systematically show cases to the user/trainee in order to give them the same skill level and experience at the end of training. To verify if the hypothesis is correct and if the CAL system has added value, a user study can be carried out.
Next to this, the CAL system might also be used as a platform to let the user/trainee practice with DL algorithms. A lot of research recently, is focussed on the development of DL algorithms to assist the radiologist in their diagnostic tasks. There has even been some speculation that these algorithms will replace radiologists in the future. However, in practice such techniques are not widely implemented into the clinic. Since the DL algorithms are used as an aid, their clinical performance is also dependent on how they are used. The CAL system also shows potential to be a platform to practice this DL aided diagnosis.
The design and development of a CAL system is not a small task. For this reason, the research phase of the Master Assignment will also be used to narrow down the scope and to assess which functionalities are most feasible for the time period and to what extent. Functionalities and steps that need to be realised:
- Decide on the type of diagnosis; classification, detection or segmentation
- Research of DL networks, e.g.:
- Which do we want to use and how many?
- How do we realise variation in performance and how much?
- Creating a systematic ordering of cases using the DL networks
- Perform a user study to validate the systematic ordering
Functionalities and steps that may be added, but need to be assessed first:
- Design of the CAL system as an application
- Creating a feedback loop in the CAL to give the user feedback during diagnosis
- Add a functionality that enables practice with DL aided diagnosis