Volumetric Segmentation and Characterisation of the Paracingulate Sulcus on MRI Scans

Junwei Yang, Duo Wang, Colleen Rollins, Matthew Leming, Píetro Lió, John Suckling, Graham K. Murray, Jane Garrison, Arnaud CachiabioRxiv 2019 · DOI

Abstract

Many architectures of deep neural networks have been designed to solve specific biomedical problems, among which segmentation is a critical step to detect and locate the boundaries of the target object from an image. In this paper, we develop a deep learning based framework to automatically segment the paracingulate sulcus (PCS) from the MRI scan and estimate lengths for its segments. The study is the first work on segmentation and characterisation of the PCS, and the model achieves a Dice score of over 0.77 on segmentation, which demonstrates its potential for clinical use to assist human annotation. Moreover, the proposed architecture as a solution can be generalised to other problems where the object has similar patterns.

Introduction

The paracingulate sulcus (PCS) is a prominent fold on the medial surface of the human brain's frontal lobe. Its morphological characteristics are of significant clinical interest; for instance, a 1 cm reduction in PCS length has been associated with a 20% increase in the risk of hallucinations in patients with schizophrenia. Despite its importance in cognitive control and mental health, the PCS is notoriously difficult to study due to its high inter-individual variability and its similarity to adjacent cortical structures.

Currently, identifying and measuring the PCS relies heavily on manual tracing by experts, a process that is time-consuming, subjective, and difficult to scale for large neuroimaging studies. This research presents the first automated deep learning framework designed to locate, segment, and characterize the PCS in 3D MRI scans.

Methodology

The researchers implemented a "coarse-to-fine" three-stage pipeline to address the challenges of segmenting a small, localized structure within the complex volume of the brain.

The proposed multi-stage pipeline: from coordinate detection and localized cropping to 3D segmentation and final length estimation.
  1. Coordinate Detection: A convolutional neural network (CNN) first predicts the center coordinate of the PCS. This step effectively "zooms in" on the region of interest, preventing the model from being confused by similar sulcal patterns in other brain regions.
  2. 3D Segmentation: Using the predicted coordinates, the MRI volume is cropped. A 3D U-Net architecture then performs voxel-wise segmentation. To handle the class imbalance (where PCS voxels are rare compared to the background), a weighted binary cross-entropy loss was employed.
  3. Length Estimation: The final stage converts the 3D segmentation mask into a clinical metric. An algorithm identifies distinct sulcal segments and calculates their length based on the distance between the furthest points in each segment.

Dataset and Training

The study utilized a dataset from the Brain Mapping Unit at the University of Cambridge, consisting of 307 T1-weighted structural MRI scans with 614 expert-annotated PCS labels. To overcome the limitations of a relatively small medical dataset, the authors utilized translation-based data augmentation, expanding the training set to over 10,000 samples.

Results and Discussion

The framework demonstrated robust performance across several metrics. The coordinate detector localized the PCS center with an average error of approximately 10mm (19 voxels). This high precision enabled the segmentation network to achieve a Dice similarity coefficient of 77.43%, nearly matching the performance achieved when using ground-truth coordinates (78.10%).

A sagittal MRI slice showing the automated segmentation of the paracingulate sulcus highlighted in white.

The automated length estimation yielded an error rate of 21.63%. While higher than the segmentation error, this figure is comparable to the inherent 16.4% measurement error found in human-generated labels when processed by the same algorithm, suggesting the model's performance is approaching the practical limits of the current ground truth.

Conclusion

This study provides a proof-of-concept for the automated morphological analysis of complex brain folds. By combining localization with high-resolution 3D segmentation, the proposed framework offers an objective and efficient alternative to manual tracing. Future work will focus on refining the length estimation algorithms through skeletonization techniques and validating the tool's predictive power in larger clinical cohorts.

Summary updated

Abstract

ABSTRACT Many architectures of deep neural networks have been designed to solve specific biomedical problems, among which segmentation is a critical step to detect and locate the boundaries of the target object from an image. In this paper, we develop a deep learning based framework to automatically segment the paracingulate sulcus (PCS) from the MRI scan and estimate lengths for its segments. The study is the first work on segmentation and characterisation of the PCS, and the model achieves a Dice score of over 0.77 on segmentation, which demonstrates its potential for clinical use to assist human annotation. Moreover, the proposed architecture as a solution can be generalised to other problems where the object has similar patterns.
BibTeX
@misc{yang2019volumetric,
  author = {Junwei Yang and Duo Wang and Colleen Rollins and Matthew Leming and Píetro Lió and John Suckling and Graham K. Murray and Jane Garrison and Arnaud Cachia},
  title = {{Volumetric Segmentation and Characterisation of the Paracingulate Sulcus on MRI Scans}},
  year = {2019},
  doi = {10.1101/859496},
  howpublished = {bioRxiv},
}
Figure