Automated Human Habenula Segmentation from T1-weighted Magnetic Resonance Images using V-Net

The Habenula (Hb), a small bilateral midbrain structure, plays an important role in aversion and reward processing. Despite increasing interest in imaging human Hb structure and function, it is challenging to define the Hb in magnetic resonance imaging (MRI) due to its small size and low anatomical contrast to the surrounding thalamus. In this study, we developed a deep learning-based human Hb segmentation model. This model was trained and tested using about a thousand subjects 3T T1-weighted images with Hb labels from our previous myelin content-based Hb segmentation. The predicted Hb segmentation showed high similarity and small surface distance to the label; the average (across test subjects) dice similarity coefficient, mean distance, and Hausdorff distance were 0.79, 0.22 mm, and 1.63 mm, respectively. We also demonstrated out-of-sample robustness using other 7T and 3T MRI datasets. The proposed Hb segmentation method is automated, objective, fast, robust, and reliable, relieving the requirement of T2-weighted images in the established myelin content-based Hb segmentation to suite a broader range of MRI studies.