Abstract
Accurately detecting and extracting major cortical sulci from brain images is vital for neuroanatomical studies, yet manual annotation is a time-consuming process governed by complex protocols. This paper proposes a supervised learning algorithm for the automated extraction of major cortical sulci from MRI volumes and cortical surfaces. By employing a Probabilistic Boosting Tree (PBT) to combine hundreds of multi-scale features, the method generates a probability map for sulcal locations. A dynamic programming approach then extracts the optimal curve based on this map and a shape prior. The algorithm is highly efficient, processing a single sulcus in under one minute, and demonstrates robust performance on both normal brains and those from patients with Williams Syndrome.
Introduction: The Challenge of Sulcal Mapping
Cortical sulci—the prominent grooves of the cerebral cortex—serve as essential landmarks for functional brain mapping and the study of neurodegenerative diseases. Despite their importance, identifying these structures is a significant bottleneck. Manual annotation is labor-intensive and requires experts to follow rigid protocols. Existing automated methods often rely on a limited set of predefined geometric rules, such as mean curvature thresholds. However, due to the high variability and noise in cortical geometry, these rule-based systems frequently struggle to distinguish major sulci from surrounding structures.
Furthermore, many automated tools require extensive preprocessing, such as tissue segmentation or canonical space mapping. Errors in these initial stages can propagate through the entire extraction process. This research introduces a flexible, data-driven approach that bypasses these limitations by learning the discriminative features of sulci directly from raw data.
Methodology: A Learning-Based Pipeline
The proposed framework moves away from rigid geometric rules toward a supervised learning model. The pipeline consists of three primary stages:
- Probabilistic Boosting Tree (PBT): Instead of relying on a few hand-picked features, the system evaluates nearly 8,000 candidate features, including 3D Haar filters, gradients, and shape indices. The PBT algorithm selects the most discriminative features to build a hierarchical classifier, producing a 3D probability map that indicates the likelihood of each voxel belonging to a specific sulcus.
- Automated End-Point Detection: The system utilizes a similar PBT-based approach to automatically identify the start and end points of a sulcus, eliminating the need for manual initialization.
- Curve Extraction via Dynamic Programming (DP): To convert the probability map into a clean anatomical curve, the authors use DP to find an optimal path that minimizes an energy function. This function balances the PBT-derived probability with a shape prior to ensure the resulting curve is smooth and anatomically plausible.
The use of "integral volumes" allows for the rapid computation of 3D Haar filters, making the entire extraction process extremely fast—typically under one minute per sulcus.
Experimental Setup and Datasets
The algorithm was evaluated using two primary datasets:
- Normal Brains: 40 MRI volumes of the right hemisphere (15 for training, 25 for testing).
- Williams Syndrome (WS) Brains: 40 MRI volumes from patients with Williams Syndrome. This dataset served as a "stress test" for the algorithm's ability to generalize to brains with abnormal and highly complex folding patterns.
- Ground Truth: All sulci were manually delineated by expert neuroanatomists to provide a gold standard for performance comparison.
Results and Discussion
The algorithm achieved high accuracy, with average distance errors ($H_{av}$) between the automated curves and manual labels ranging from 3 to 5 mm. Given that the typical inter-expert variability is approximately 2 mm, the algorithm's performance approaches human-level consistency.
The system also demonstrated remarkable robustness when applied to the Williams Syndrome dataset. While the error was slightly higher (e.g., 4.72 mm vs. 3.42 mm for the Central Sulcus), the results confirm that the learning-based approach is flexible enough to handle pathological anatomy. The speed of the algorithm—running in under a minute on a standard PC—represents a significant advancement over manual methods that can take hours.
Conclusion and Future Directions
By learning the specific "look" of a sulcus in 3D space and utilizing efficient computational techniques like integral volumes, this framework provides a fast and robust solution for cortical mapping. It successfully bypasses the need for error-prone preprocessing steps like tissue segmentation. Future work will focus on incorporating joint statistical regularities between different sulci to further improve accuracy by teaching the model how these structures relate to one another spatially.
Summary updated