TL;DR
This paper introduces PSACNN, a novel Convolutional Neural Network (CNN) approach for fast and accurate brain segmentation in MRI scans. It tackles the challenge of variations in MRI acquisition protocols by using "pulse sequence augmentation" to train a CNN that is robust to these differences, achieving state-of-the-art accuracy with significantly reduced processing time.
What Problem Does This Paper Address?
Accurate segmentation of brain anatomy from T1-weighted MRI scans is a crucial but computationally intensive step in neuroimaging pipelines. While deep learning methods like CNNs have dramatically sped up this process, their supervised nature makes them vulnerable to variations in MRI acquisition protocols (e.g., different scanners, field strengths, pulse sequences). These variations, common in multi-center studies, can introduce biases and reduce segmentation accuracy. Existing methods struggle to generalize across these diverse acquisition settings without extensive retraining or manual intervention. This paper aims to develop a segmentation algorithm that is both highly accurate and fast, yet also resilient to the inherent variability in MRI data.
Methods
The core innovation is the Pulse Sequence Augmented Convolutional Neural Network (PSACNN). The approach involves a three-step training process:
- Estimating Test Data Acquisition Parameters: The algorithm first estimates the pulse sequence parameters of the test MRI data. This is done by creating approximate forward models for common pulse sequences like MPRAGE and FLASH, which relate image intensity to tissue properties (proton density, T1, T2 relaxation times) and acquisition parameters (TR, TE, flip angle). These models allow the estimation of key parameters directly from the image intensities.
- Feature Extraction and Augmentation: For training, the authors use a U-Net CNN architecture. To handle memory constraints, they work with 32x32x32 patches. Crucially, they augment the training data by generating synthetic MRI patches that mimic the characteristics of the test data's pulse sequence. This is achieved by using the estimated test data parameters and the forward models to transform NMR parameter maps from the training data into simulated images that resemble the test data's contrast. They also incorporate warped FreeSurfer atlas coordinates to provide global context.
- Training a Deep CNN: The augmented and original training data are used to train the U-Net. The network learns to predict segmentation labels from these enhanced feature representations.
Datasets
The paper utilizes existing datasets for training and evaluation. The primary training dataset consists of T1-weighted MRI images and their corresponding expert manual segmentations. For evaluation, they use:
- Siemens Dataset: 14 subjects with MPRAGE scans from a Siemens SONATA scanner.
- GE Dataset: 13 subjects with SPGR scans from a GE Signa scanner.
- Multi-scanner Consistency Datasets: Four datasets from 15 subjects, including MEF (multi-echo FLASH on Siemens 3T), TRIO (MPRAGE on Siemens TRIO), GE (MPRAGE on 1.5T GE Signa), and Siemens (MPRAGE on 1.5T Siemens SONATA).
The authors also mention generating synthetic T1-w images using image synthesis from multi-echo FLASH data to create NMR parameter maps (p, T1, T2) for their atlas dataset, which is a common practice but here specifically used to complement existing atlas sets.
Key Insights and Ideas
- Pulse Sequence Augmentation is Key: The central idea is that by simulating the target acquisition protocol during training, the CNN can learn to be invariant to variations in pulse sequences. This is a powerful way to bridge the gap between training data and real-world, diverse test data.
- Approximate Forward Models are Sufficient: The authors demonstrate that highly accurate, albeit approximate, forward models of MRI physics are sufficient to generate effective training augmentations. This avoids the need for complex, full Bloch equation simulations for every possible pulse sequence.
- Combining Intensity and Spatial Information: The approach effectively combines local image intensity information with global spatial context derived from atlas-based warping, leading to more robust segmentations.
- Speed and Accuracy Trade-off: PSACNN achieves state-of-the-art accuracy while being orders of magnitude faster than traditional methods like multi-atlas segmentation, making it practical for large-scale studies.
Key Results
- High Accuracy: PSACNN achieves state-of-the-art segmentation accuracy, with an overall Dice overlap of 0.94 reported in the abstract. In experiments, it significantly outperforms unaugmented CNNs and other methods like FreeSurfer, SAMSEG, and MALF on datasets with varying acquisition protocols.
- Robustness to Protocol Variations: On the GE dataset, which presents a noticeably different tissue contrast, PSACNN maintained comparable accuracy to state-of-the-art algorithms, whereas an unaugmented CNN performed poorly.
- Speed: The algorithm produces segmentations within seconds to minutes, a significant improvement over traditional methods that can take hours.
- Multi-scanner Consistency: PSACNN shows good consistency across different scanners and protocols, outperforming other methods in several scenarios.
Limitations and Future Work
- Limitations:
- The paper acknowledges that it is generally not possible to robustly estimate all pulse sequence parameters purely from image intensities due to their highly nonlinear dependence.
- While the approximate forward models fit well for certain T1 ranges, their accuracy for extreme values might be a limitation.
- The current implementation works with 32x32x32 patches due to GPU memory constraints, which might limit the capture of very large structures in a single patch.
- Future Work:
- The authors intend to use more accurate imaging equations and simulate other pulse sequences to expand the range of applications.
- Further work could explore simulating a wider variety of scanner parameters beyond just pulse sequence variations.
Summary updated