A Self-supervised Deep Learning Model for Diagonal Sulcus Detection with Limited Labeled Data

Sulci are a fundamental part of brain morphology, closely linked to brain function, cognition, and behavior. Tertiary sulci, characterized as the shallowest and smallest subtype, pose a challenging task for detection. The diagonal sulcus (ds), located in a crucial area in language processing, has a prevalence between 50% and 60%. Automatic detection of the ds is an unexplored field: while some sulci segmenters include the ds, their accuracy is usually low. In this work, we present a deep learning based model for ds detection using a fine-tuning approach with limited training labeled data. A convolutional autoencoder was employed to learn specific features related to brain morphology with unlabeled data through self-supervised learning. Subsequently, the pre-trained network was fine-tuned to detect the ds using a less extensive labeled dataset. We achieved a mean F1-score of 0.7176 (SD=0.0736) for the test set and a F1-score of 0.72 for a second held-out set, surpassing the results of a standard software and other alternative deep learning models. We conducted an interpretability analysis of the results using occlusion maps and observed that the models focused on adjacent sulci to the ds for prediction, consistent with the approach taken by experts in manual annotation. We also analyzed the challenges of manual labeling by conducting a thorough examination of interrater agreement on a small dataset and its relationship with our model's performance. Finally, we applied our method on a population analysis and reported the prevalence of ds in a case study.