A benchmark for hypothalamus segmentation on T1-weighted MR images

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Computational Neuroscience

Background:

• The hypothalamus is crucial for homeostasis, but manual segmentation in MRIs is subjective and variable.
• Existing automated hypothalamus segmentation methods lack generalization across different datasets.
• Accurate hypothalamic volumetry is vital for understanding neuropsychiatric disorders like schizophrenia and Alzheimer's disease.

Purpose of the Study:

• To develop a robust, fully automated method for hypothalamus segmentation on T1-weighted MR images.
• To create a diverse benchmark dataset for evaluating hypothalamus segmentation models.
• To improve the generalization ability of automated hypothalamus segmentation.

Main Methods:

• A benchmark dataset was created with 1381 subjects from IXI, CC359, OASIS, and MiLI datasets, including manual and automatic annotations.
• A novel teacher-student-based model with segmentation and correction blocks was developed for automated hypothalamus segmentation.
• The model was trained on MiLI, IXI, and CC359 datasets and tested on unseen data.

Main Results:

• The proposed model achieved a Dice coefficient of 0.83 on a test set comprising IXI, CC359, and MiLI data.
• On the OASIS dataset (not used during training), the model achieved a Dice coefficient of 0.74, demonstrating good generalization.
• The dataset, baseline model, and code are publicly available to foster further research.

Conclusions:

• The developed HypAST benchmark and teacher-student model offer a significant advancement in automated hypothalamus segmentation.
• The method demonstrates improved generalization capabilities compared to existing approaches.
• This work provides a valuable resource for researchers studying the hypothalamus in various neurological and psychiatric conditions.