TL;DR
This paper evaluates three leading methods for automatically identifying the brain's mid-sagittal plane, a crucial step in many neuroimaging analyses. The study finds that these automated methods perform comparably to human experts, with significant errors only occurring in cases of cerebral atrophy.
What Problem Does This Paper Address?
Accurately identifying the mid-sagittal plane (MSP), the imaginary line that divides the left and right hemispheres of the brain, is fundamental for various neuroimaging applications. These include analyzing asymmetries between hemispheres, correcting image misalignments, and serving as a starting point for image registration. While numerous methods have been proposed over the years, many lack rigorous, comparative validation against each other and against human performance. This paper aims to fill that gap by classifying existing methods and evaluating representative examples from each class on a standardized dataset, comparing their accuracy and speed to manual delineations by experienced observers.
Methods
The authors categorize existing MSP extraction methods into three main approaches:
- Feature-based methods: These rely on specific visual cues of the interhemispheric fissure (IF), such as its linear shape or low intensity due to cerebrospinal fluid (CSF).
- Global symmetry-based methods: These assume bilateral symmetry of the brain hemispheres and work by reflecting the entire image and then rigidly registering it to the original. The transformation derived from this registration helps define the MSP.
- Local symmetry-based methods: These first roughly locate the brain's center and then optimize a symmetry measure within a localized region around the expected IF. This approach is generally faster as it focuses on a smaller area.
The study implemented one leading method from each category: a feature-based method (Volkau et al./Nowinski et al.), a local symmetry-based method (Hu et al.), and a global symmetry-based method (using elastix with mutual information for registration).
Datasets
The study evaluated the implemented methods on MRI data from 50 participants in the Second Manifestations of ARTerial disease (SMART) study. These participants had a mean age of 60 years and a standard deviation of 9 years. The data consisted of transversal T1-weighted gradient echo sequences with a voxel size of 0.9 x 0.9 x 4.0 mm. A brain mask was generated using the BET tool for evaluation. Manual delineations were performed by two experienced observers, with observer A performing a second delineation to assess intra-observer variability.
Key Insights and Ideas
- Human-level performance is achievable: The most striking finding is that automated methods can achieve accuracy comparable to that of experienced human observers in identifying the MSP. This suggests that current automated techniques are mature enough for many clinical and research applications.
- Cerebral atrophy is a major challenge: The study highlights that participants with cerebral atrophy, characterized by a widened IF, pose the greatest difficulty for both automated methods and human observers. This widening allows for multiple plane orientations to appear "correct," leading to larger errors.
- Speed vs. Accuracy trade-off: While all methods performed similarly in terms of accuracy, there was a significant difference in computation time. Local symmetry-based methods were the fastest, followed by feature-based methods, with global symmetry-based methods being considerably slower due to the full 3D registration.
- Beyond a simple plane: The paper acknowledges that the IF is not always a perfect geometric plane, especially in cases of "brain torque" (where lobes are asymmetrically sized). This suggests that for some applications, a curved surface might be a more appropriate representation than a flat plane.
Key Results
- Accuracy: No significant differences were found between the errors (angular and volume) made by the three automated methods and the inter-observer differences. All automated methods and inter-observer differences showed significantly larger errors than intra-observer differences.
- Feature-based: Angular Error (AE) 1.25° (0.53°), Volume Error (VE) 1.58 ml (0.59 ml)
- Global symmetry-based: AE 1.54° (0.61°), VE 1.67 ml (0.54 ml)
- Local symmetry-based: AE 1.46° (0.60°), VE 1.68 ml (0.62 ml)
- Inter-observer: AE 1.39° (0.95°), VE 1.60 ml (0.87 ml)
- Intra-observer: AE 1.02° (0.66°), VE 1.08 ml (0.55 ml)
- Speed:
- Local symmetry-based: 0.5 seconds
- Feature-based: 1.8 seconds
- Global symmetry-based: 33.6 seconds
Limitations and Future Work
- Cerebral Atrophy: The study identifies cerebral atrophy as a significant limitation, leading to larger errors. Future work could explore methods specifically designed to handle widened IFs or investigate how to better segment brain tissue in such cases.
- Curved Surfaces: For cases with significant "brain torque," a curved mid-sagittal surface might provide a better separation than a plane. The authors suggest this as a direction for future research.
- Generalizability: While the methods are expected to work on other MRI sequences (e.g., FLAIR, IR, T2) and modalities (e.g., CT), higher spatial resolution images might increase computation times.
- Method Implementation: The authors note that while they implemented leading methods, slight differences from the original published versions might exist. For global symmetry-based methods, a generic implementation was used, and task-specific improvements could potentially reduce errors.
Summary updated