← Back to Previous Page

 

S041: A MULTIMODAL SCORE-BASED PIPELINE FOR AUTOMATED PITUITARY SEGMENTATION IN MRI
Charbel Marche, MS¹; John Van Horn, PhD²; Michael Catalino³; ¹University of Virginia School of Medicine; ²University of Virginia School of Data Science; ³University of Virginia Department of Neurosurgery

Accurate segmentation of the pituitary gland is essential for studying normal anatomy and for developing automated tools to detect adenomas and invasion of surrounding structures. Manual segmentation is time-intensive and prone to variability, motivating reproducible automatic approaches.

We developed and validated a tunable pituitary segmentation pipeline that leverages paired T1- and T2-weighted MRIs from the Human Connectome Project’s dataset of 100 unrelated subjects, registered into Montreal Neurological Institute (MNI) space. Preprocessing included motion correction, noise smoothing, linear T1-to-T2 registration, and subsequent linear, then nonlinear registration into MNI coordinates. Ground-truth pituitary masks were manually generated in 3D Slicer for 35 random subjects (19 female, 16 male) using the MNI-registered images. Within a tuned, pituitary-centered region of interest, voxels were scored using a composite framework incorporating (i) distance from the centroid of the pituitary in normal space, (ii) modality-specific intensity ranges, and (iii) connectivity to high-confidence neighbors. Scores were combined into a weighted final score, with thresholds optimized during training. Appendage removal and morphological smoothing were optionally applied to refine boundaries using scikit-image operations.

For tuning and evaluation, subjects were divided into training (70%) and testing (30%) sets separately for males and females. A total of 19 configuration variables with 23 tunable parameters were optimized across 600 Optuna experiments (300 for female and 300 for male) to maximize the mean Dice coefficient penalized by variance, ensuring both accuracy and consistency across subjects.  The method achieved a mean Dice coefficient of 0.80 across all subjects. Gender-stratified analysis demonstrated improved performance in selected females (0.82 ± 0.02) compared to selected males (0.79 ± 0.05). Testing on held-out cases yielded consistent results (0.80 ± 0.03 in females; 0.80 ± 0.04 in males).

We present an automatic, interpretable pituitary segmentation pipeline that achieves human-level performance and scales to large MRI datasets. By providing a transparent set of parameters for scoring voxels as pituitary tissue, this framework lays the groundwork for automated detection of invasive pituitary adenomas and abnormal tissue in future applications.

Link to source code: https://github.com/Pituitary-Normal-Space/pituitary-segmentation-tool

 

← Back to Previous Page