TL;DR
This paper introduces a novel deep learning method, a 3D U-Net, for automatically segmenting both white matter (WM) and cortical lesions (CLs) in multiple sclerosis (MS) patients using only two common MRI sequences (FLAIR and MP2RAGE). This approach significantly improves the detection of challenging-to-see cortical lesions, offering a more efficient and accurate tool for MS research and potentially clinical practice.
What Problem Does This Paper Address?
Multiple Sclerosis (MS) is a chronic neurological disease characterized by lesions in the brain. While white matter lesions (WMLs) have been extensively studied and are relatively easy to detect with standard MRI, cortical lesions (CLs) are increasingly recognized as crucial biomarkers, especially for cognitive and physical disability. However, CLs are much harder to spot on conventional 3T MRI scans due to their subtle appearance. This makes manual segmentation time-consuming and prone to variability, hindering large-scale studies and clinical application. Existing automated methods often struggle with CLs, either requiring specialized MRI sequences that aren't widely used or failing to detect smaller lesions. This paper aims to bridge this gap by developing an automated segmentation method that is accurate for both WMLs and CLs, using a clinically practical set of MRI sequences.
Methods
The core of this work is a 3D U-Net architecture, a type of convolutional neural network (CNN) well-suited for image segmentation. The authors adapted the standard U-Net by reducing the number of resolution levels to three, resulting in a more compact model with fewer trainable parameters (3.8 million) to mitigate overfitting, especially given the challenge of segmenting small CLs. This 3D architecture is crucial for leveraging the volumetric information present in the MRI data.
Key methodological aspects include:
- Input Data: The model is trained and tested using only two MRI contrasts: FLAIR (a conventional sequence good for WMLs) and MP2RAGE (a specialized sequence that enhances CL visibility). This is a significant advantage as it uses a clinically plausible and less demanding acquisition protocol compared to methods requiring additional sequences like DIR.
- Pre-processing: A simple rigid registration based on mutual information is performed to align the FLAIR and MP2RAGE images. The authors deliberately avoided complex pre-processing steps like skull stripping to reduce variability and focus on the network's inherent capabilities.
- Training Strategy: A weighted cross-entropy loss function is employed, giving CLs a higher weight (5x) than WMLs and background to address the class imbalance. Data augmentation (rotation, scaling, flipping) is applied on-the-fly to improve robustness.
- Evaluation: The method is evaluated using two scenarios: a 6-fold cross-validation across all subjects and an independent test set to assess generalization. Standard metrics like Dice coefficient (DSC), Absolute Volume Difference (AVD), Positive Predicted Value (PPV), True Positive Rate (LTPR), and False Positive Rate (LFPR) are used.
Datasets
The study utilizes two datasets comprising a total of 90 MS patients:
- Dataset I: Acquired at Basel University Hospital, with 54 patients at various disease stages.
- Dataset II: Acquired at Lausanne University Hospital, with 36 patients in the early stages of MS.
These datasets were manually segmented by experienced neurologists and radiologists, resulting in a total of 3856 WMLs and 728 CLs. The paper notes that CLs were further classified into subtypes (leukocortical, intracortical, subpial), with leukocortical lesions being the most prevalent. A minimum lesion size of 3 voxels was enforced for both ground truth and predictions.
Key Insights and Ideas
- MP2RAGE is Key for CLs: The study strongly reinforces that specialized sequences like MP2RAGE are essential for improving the detection of cortical lesions, offering significantly better contrast than FLAIR alone.
- 3D U-Net Simplicity and Power: The authors demonstrate that a relatively simple 3D U-Net architecture, without extensive feature engineering or complex pre-processing, can achieve state-of-the-art performance for both WML and CL segmentation.
- Clinically Plausible Input: By using only FLAIR and MP2RAGE, the proposed method is more readily adoptable into clinical workflows compared to methods requiring more specialized or time-consuming sequences.
- Generalization Across Scanners: The method shows promising generalization capabilities, performing well on data acquired from different scanners at different hospitals, which is crucial for real-world applicability.
- Small Lesion Detection is Challenging: While the method performs well overall, the detection of very small lesions (3-10 voxels) remains a challenge, likely due to partial volume effects and artefacts.
Key Results
- The proposed 3D U-Net achieved a median 76% detection rate for both cortical and white matter lesions with a median 29% false positive rate across all subjects in the cross-validation.
- In the cross-validation, the 3D U-Net significantly outperformed two established methods (LST-LGA and nicMSlesions) in CL and WML detection rate and false positive rate.
- The method achieved a 71% CL detection rate on the independent test set, outperforming other methods in most metrics, including a strong performance in WM LTPR.
- Inference time for the 3D U-Net was remarkably fast, around 20 seconds per subject (excluding registration), significantly faster than the other methods.
- The Pearson correlation coefficient between manual and automatically segmented lesion volumes was high, reaching r=0.94 for the 3D U-Net.
Limitations and Future Work
- Small Lesion Detection: The authors acknowledge that detecting very small lesions (3-10 voxels) is still a challenge and suggest that including more such lesions in the training data could improve performance.
- CL Subtype Classification: Currently, the method segments all CLs together. Future work aims to extend the output to classify lesions into different subtypes (leukocortical, intracortical, subpial), which would add significant clinical value.
- Inter- and Intra-Rater Reliability: The study did not compare the automatic segmentation to inter- and intra-rater reliability, as ground truth masks were delineated only once by experts.
- CL Type Limitation: The majority of CLs in the dataset were leukocortical/juxtacortical, suggesting the method's current accuracy for other CL types might be limited.
- Future Directions: The authors suggest exploring the T1 map acquired with MP2RAGE for improved lesion delineation and developing methods for classifying lesion types. They also highlight the potential for reducing MP2RAGE acquisition time, making it even more clinically viable.
Summary updated