Deep learning methods on computed aided diagnosis tasks have become extremely
performant and often outperform radiologists. They are however (pre)trained on massive amounts of
data. On the other hand, the radiologist has learnt a set of very clear indicators for malignancy and
just knows that certain shapes, sizes, locations etc correlate with malignancy. Thanks to the large
amounts of data fed into the neural networks, the system is expected to learn these patterns. It would
however be conceptually nice that these cues could just be taught to the network.
One way to do this, is to combines simple probabilistic rules (if the lesion is larger than x, then the
probability of malignancy is high if the lesion is round, then the probability of malignancy is small …)
with deep learning methods to detect these properties. A nice framework seems to be deepproblog
[1] but there exist no studies evaluating this in a medical context.

To implement and compare several pipelines for image-based cad systems with and
without neuro symbolic AI. Be the first to show that this idea works on computer aided diagnosis
systems (CAD).

Literature Review (ETOC: 2 months): Familiarize with existing literature on deepproblog,
get the basic demos working and reproduce their results.
- Identify the relevant image features and proper deep learning methods for computing these
features.
- Integrate domain knowledge in the DL pipeline via deepproblog.
- Comparative analysis.

(Mandatory) qualifications:
• Following an MSc in a field related to one or more of the following: Computer
Science, Biomedical Engineering, Applied Computer Science - Digital Health.
• Strong programming skills (Python).
• Ability to write scientific reports and communicate research results at
conferences in English.