A Learning Based Algorithm for Automatic Extraction of the Cortical Sulci

Songfeng Zheng, Zhuowen Tu, Alan L. Yuille, Allan L. Reiss, Rebecca A. Dutton, Agatha D. Lee, Albert M. Galaburda, Paul M. Thompson, Ivo D. Dinov, Arthur W. Toga

Abstract

This paper presents a fully automatic, learning-based framework for extracting major cortical sulci directly from 3D MRI volumes. By integrating a Probabilistic Boosting Tree (PBT) for voxel classification with Dynamic Programming (DP) for curve tracing, the method eliminates the need for manual initialization or complex tissue segmentation, providing a robust tool for large-scale neuroimaging studies.

Introduction

Cortical sulci—the "valleys" of the brain's folded surface—serve as critical landmarks for neuroimaging, brain mapping, and the study of neurodegenerative diseases like Alzheimer’s. However, manual extraction is a labor-intensive process that requires significant anatomical expertise. Existing automated solutions often come with "strings attached," such as requiring manual seed points or extensive preprocessing like brain surface mapping and tissue segmentation. These steps introduce potential for error and slow down research pipelines.

This research proposes a "plug-and-play" algorithm designed to operate directly on raw MRI data with minimal human intervention, bypassing traditional bottlenecks in neuroimaging workflows.

Methodology: A Two-Step Optimization

The extraction process is framed as a two-step optimization problem: discriminative modeling followed by curve extraction.

1. Learning "Sulcus-ness" with PBT

Rather than relying on hard-coded geometric rules (such as simple curvature measures), the algorithm learns to identify sulci from a massive pool of approximately 5,000 candidate features. These include intensity, gradients, and 3D Haar filters. A Probabilistic Boosting Tree (PBT) is employed to build a hierarchical classifier that discovers the complex photometric and geometric patterns experts use intuitively. For each voxel, the PBT generates a probability map indicating the likelihood of it belonging to a major sulcus.

The Probabilistic Boosting Tree (PBT) hierarchical classification framework.

To maintain computational efficiency while processing thousands of features, the authors utilize integral volumes. This allows 3D Haar filter responses to be calculated in constant time, regardless of the filter's size.

2. Curve Extraction via Dynamic Programming

Because probability maps can be noisy or disconnected, the framework uses Dynamic Programming (DP) to convert these maps into continuous 3D curves. An energy function balances the PBT-derived likelihood with a "shape prior" that discourages unrealistic, jagged paths. The system also features an adaptive method to automatically select start and end points based on training data statistics, ensuring the process is truly automated.

Experimental Results

The algorithm was evaluated on a dataset of 40 MRI volumes, with 15 used for training and 25 for testing. Several major sulci, including the Central and Middle Frontal sulci, were manually labeled by experts to serve as ground truth.

Visual comparison of ground truth, PBT probability maps, and final extraction results across different sulci and data formats.

Performance and Accuracy

  • Accuracy: The average distance between the automatically extracted curves and expert labels was between 3 to 5 voxels.
  • Generalization: Testing errors were only slightly higher than training errors, indicating the model generalizes well to unseen brain scans.
  • Efficiency: On 2006-era hardware (2.4 GHz CPU), the PBT classification took approximately 60 seconds, and the DP extraction took 20 seconds. This total processing time of under 90 seconds is significantly faster than manual tracing.
  • Versatility: The method demonstrated equal efficacy on both raw 3D volumes and 2D cortical surfaces, proving the learning-based approach is not tied to a specific data format.

Conclusion

The proposed framework successfully automates cortical sulcus extraction without the need for tedious preprocessing. While endpoint estimation remains a primary source of error—largely due to inherent anatomical ambiguity that even human experts face—the framework's speed and accuracy make it a viable solution for modern neuroimaging. Future work will focus on refining endpoint precision and applying this learning-based framework to other object detection tasks within medical imaging.

Summary updated

Abstract

This paper presents a learning based method for automatic extraction of the major cortical sulci from MRI volumes or extracted surfaces. Instead of using a few pre-defined rules such as the mean curvature properties, to detect the major sulci, the algorithm learns a discriminative model by selecting and combining features from a large pool of candidates. We used the Probabilistic Boosting Tree algorithm [16] to learn the model, which implicitly discovers and combines rules based on manually annotated sulci traced by neuroanatomists. The algorithm almost has no parameters to tune and is fast because of the adoption of integral volume and 3D Haar filters. For a given approximately registered MRI volume, the algorithm computes the probability of how likely it is that each voxel lies on a major sulcus curve. Dynamic programming is then applied to extract the curve based on the probability map and a shape prior. Because the algorithm can be applied to MRI volumes directly, there is no need to perform preprocessing such as tissue segmentation or mapping to a canonical space. The learning aspect makes the approach flexible and it also works on extracted cortical surfaces.
BibTeX
@inproceedings{zheng2006learning,
  author = {Songfeng Zheng and Zhuowen Tu and Alan L. Yuille and Allan L. Reiss and Rebecca A. Dutton and Agatha D. Lee and Albert M. Galaburda and Paul M. Thompson and Ivo D. Dinov and Arthur W. Toga},
  title = {{A Learning Based Algorithm for Automatic Extraction of the Cortical Sulci}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2006},
  doi = {10.1007/11866565\_85},
}
Figure