TL;DR
This paper introduces a novel two-step method for segmenting sub-acute ischemic stroke lesions from FLAIR MRI scans, utilizing Gabor texture features and a customized Markov Random Field (MRF). The approach achieves competitive accuracy using only FLAIR data, outperforming recent deep learning and level-set methods in head-to-head comparisons.
What Problem Does This Paper Address?
Accurately measuring the volume of stroke lesions is increasingly important for clinical trials, serving as a more sensitive endpoint than traditional clinical scores. However, automatically segmenting these lesions from medical images, particularly FLAIR MRI sequences commonly used in follow-up scans, is challenging. Existing methods often struggle with variations in lesion shape, size, and location, as well as image artifacts common in multi-center datasets. A key difficulty is distinguishing actual stroke lesions from other white matter abnormalities, which can lead to over-segmentation. This paper aims to develop a robust and reliable method for stroke lesion segmentation using only FLAIR MRI, addressing the need for a standardized approach that can be applied across different imaging centers and scanners.
Methods
The proposed method employs a two-stage approach:
- Initial Bayesian Classification with Gabor Textures:
- FLAIR images are processed using multi-window 2D Gabor transforms to extract texture features at each voxel. Gabor filters are chosen for their ability to capture local image characteristics at different orientations and scales.
- A Bayesian classifier then estimates the probability of each voxel belonging to either the "lesion" or "non-lesion" class based on these Gabor texture features.
- The Gabor filter parameters are optimized to maximize the separation between lesion and non-lesion classes, using a graph cut approach.
- Customized Markov Random Field (MRF) Refinement:
- The initial classification is refined using a customized MRF model. This MRF goes beyond standard intensity-based models by incorporating:
- Intensity: The signal intensity of each voxel.
- Texture: The Gabor texture features extracted in the first step.
- Intensity Differences: The differences in intensity between a voxel and its neighbors, helping to preserve lesion boundaries.
- The MRF formulation considers various clique sizes (2, 3, and 4 voxels) to capture local spatial relationships.
- Gaussian distributions are used to model the likelihoods of tissue classes, intensity differences, and texture features. The model is trained using data from the ISLES challenge and other databases.
- The final segmentation is obtained by minimizing the MRF energy function using an iterated conditional modes (ICM) approach.
Datasets
The study utilized three multi-center FLAIR MRI datasets for development and evaluation:
- SFB Database: 23 FLAIR MRI datasets from the University Medical Center Hamburg-Eppendorf, Germany. Acquired on a 3T Siemens Skyra scanner.
- I-KNOW Database: 108 FLAIR datasets from a prospective European multi-center stroke imaging study, acquired across five centers with varying 1.5T scanners and parameters.
- ISLES Database: 28 FLAIR MRI datasets publicly available from the Ischemic Stroke Lesion Segmentation (ISLES) challenge. These were acquired on 3T MRI scanners but were registered and resampled to T1-weighted images, potentially introducing interpolation artifacts.
In total, 151 FLAIR datasets were used. A leave-one-patient-out cross-validation strategy was employed. Gabor parameters were optimized using datasets from one database to ensure unbiased evaluation on another.
Key Insights and Ideas
- Texture as a Key Differentiator: The paper highlights the crucial role of Gabor texture features in distinguishing stroke lesions from other white matter abnormalities, which is a significant challenge when using FLAIR MRI alone.
- Customized MRF for Boundary Preservation: The MRF model's ability to incorporate intensity differences alongside intensity and texture is key to refining segmentation and preserving accurate lesion boundaries.
- FLAIR-Only Robustness: Demonstrating competitive performance using only FLAIR MRI is significant, as FLAIR is a standard sequence in follow-up stroke imaging, making the method broadly applicable even when multi-modal data is unavailable.
- Generalizability Challenges: The study acknowledges that while the method shows promise, variations in acquisition parameters across multi-center datasets can impact performance, necessitating careful preprocessing and potentially further optimization for specific scanner types.
Key Results
- Overall Accuracy: The proposed method achieved an average Dice coefficient of 0.582 across all 151 datasets. This is competitive with state-of-the-art methods that often utilize multi-modal imaging.
- Comparison to State-of-the-Art: The method outperformed two leading methods from the ISLES 2015 challenge: a convolutional neural network (CNN) approach (Dice 0.541) and a fuzzy C-means clustering with level-set method (Dice 0.524).
- Small Lesion Segmentation: The method successfully segmented 258 out of 483 small lesion components (< 1 ml) as true positives, with 81 false positives, showing good performance in detecting these challenging cases.
- Volumetric Analysis: While generally showing strong correlation (r=0.81) between manual and automatic lesion volumes, the method tended to underestimate larger lesions, particularly at the borders. This led to a lower Dice score for the ISLES dataset, which had pre-processed, resampled images.
- Computational Cost: Lesion segmentation took 20-30 minutes per dataset on standard datasets and 70-90 minutes on the interpolated ISLES datasets. Training the Gabor parameters was the most computationally intensive step, requiring several hours, but this is a one-time cost.
Limitations and Future Work
- Volumetric Underestimation: The paper acknowledges that the method tends to underestimate lesion volumes, especially for larger lesions, which limits its immediate use for lesion volume as a clinical endpoint.
- Generalizability: While tested across multiple datasets, the method's performance varied, particularly for the ISLES dataset due to pre-processing artifacts. A more independent validation set would be beneficial.
- Parameter Optimization: The comparison methods (CNN and level-set) were not fully re-optimized, suggesting their performance could potentially be improved.
- Future Directions: The authors suggest exploring non-uniform weighting of contrasts and textures based on voxel location to better handle lesion boundaries. They also note that integrating multi-modal MRI data is a potential avenue for further improvement.
Summary updated