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

Area of Science:

• Neuroscience
• Medical Imaging
• Artificial Intelligence

Background:

• Sulci are key brain structures influencing cognition and behavior.
• Tertiary sulci, like the diagonal sulcus (ds), are difficult to detect automatically.
• The ds is vital for language processing, with a prevalence of 50-60%.

Purpose of the Study:

• To develop and validate a deep learning model for accurate automatic detection of the diagonal sulcus (ds).
• To address the limitations of existing sulci segmentation tools in ds identification.
• To explore the application of self-supervised and fine-tuning learning for this task with limited labeled data.

Main Methods:

• Utilized a convolutional autoencoder for self-supervised learning on unlabeled brain data.
• Fine-tuned the pre-trained model on a limited labeled dataset for ds detection.
• Employed occlusion maps for model interpretability and analyzed interrater reliability in manual labeling.

Main Results:

• Achieved a mean F1-score of 0.7176 (SD=0.0736) on the test set and 0.72 on a held-out set.
• The model's performance surpassed standard software and alternative deep learning approaches.
• Interpretability analysis showed the model focuses on adjacent sulci, similar to expert annotators.

Conclusions:

• The proposed deep learning model offers a robust and accurate method for automatic diagonal sulcus detection.
• The fine-tuning approach effectively utilizes limited labeled data for specialized brain structure identification.
• The method shows promise for population-level studies investigating ds prevalence and its clinical significance.