Self-Supervised Learning for Functional Brain Networks identification in fMRI from Healthy to Unhealthy Patients

Resting State Functional Magnetic Resonance Imaging (rs-fMRI) technique is gaining more attention among medical practitioners because, it allows recognition of functional brain networks and is very suitable for complex situations where the participation of the patients is not required. This approach is also interesting for non-invasive medical imaging where healthy subjects can be enrolled very easily during the data acquisition process. However, one of its limitations is that the clinicians must manually annotate the image data. While no clinical use of this annotation is needed at any stage of neurosurgical procedure, this process is often time consuming and can only be carried our by domain experts. We investigate the possibility to perform self-supervision from healthy subject data without the need of image annotation, followed by transfer learning from the models trained on some pretext task. The result of self-supervision is shown to bring about 3% increase in performance without the effort and time of manual annotation of fMRI data by expert.