TL;DR
This paper introduces a novel, fully automatic deep learning framework for segmenting acute ischemic stroke lesions in Diffusion-Weighted Imaging (DWI) MRI scans. By combining an ensemble of DeconvNets (EDD Net) for initial lesion detection and a specialized evaluation network (MUSCLE Net) to filter false positives, the system achieves state-of-the-art performance on a large clinical dataset, significantly improving upon previous methods.
What Problem Does This Paper Address?
Stroke is a leading cause of long-term disability and death, and early diagnosis and treatment of acute ischemic lesions are crucial for improving patient outcomes. Diffusion-weighted MRI (DWI) is highly sensitive to these lesions, but manually segmenting them is time-consuming and requires specialized expertise. Existing automated or semi-automated methods often struggle with the variability in lesion size, shape, and location, and many are trained on small datasets, limiting their generalizability. Furthermore, artifacts in DWI scans can be easily mistaken for lesions, leading to inaccurate segmentations. This paper aims to bridge this gap by developing a robust, fully automatic system capable of accurately segmenting acute ischemic lesions in DWI, even in the presence of artifacts and diverse lesion characteristics.
Methods
The proposed framework consists of two main components:
- EDD Net (Ensemble of DeconvNets): This network is designed for initial lesion segmentation. It utilizes an ensemble of two DeconvNets (a type of convolutional neural network known for its semantic segmentation capabilities). The ensemble approach leverages the diversity of individual networks to improve robustness. The EDD Net takes image patches as input and outputs a probabilistic segmentation map.
- MUSCLE Net (Multi-Scale Convolutional Label Evaluation Net): This network acts as a post-processing step to refine the EDD Net's output. Its primary role is to identify and remove false positives that the EDD Net might have detected. The MUSCLE Net evaluates candidate lesions at multiple scales, focusing on smaller regions where artifacts are more likely to be confused with actual lesions. It uses the probabilistic output from the EDD Net and the original DWI image patches to make its evaluation.
The overall system is trained end-to-end for the EDD Net, while the MUSCLE Net is trained separately to evaluate candidate lesions. The use of image patches as input, rather than entire slices, is a key strategy to handle the class imbalance between normal tissue and lesions and to improve training efficiency.
Datasets
The study utilized a large dataset of Diffusion-Weighted Imaging (DWI) scans collected from 741 acute stroke patients from local hospitals. The data was collected retrospectively and anonymized.
- Training and Validation: 380 subjects were used for training and validation. Specifically, 274 subjects were used for training and 106 for validation.
- Testing: The remaining 361 subjects were used for testing. The paper notes that the data was acquired from three different Siemens scanners, implying a need for robust pre-processing to handle variations in acquisition parameters.
Key Insights and Ideas
- Ensemble Learning for Robustness: Combining multiple DeconvNets (EDD Net) significantly improves performance by leveraging their diverse error patterns, particularly in reducing false positives.
- Two-Stage Approach for False Positive Reduction: Separating lesion detection (EDD Net) from false positive evaluation (MUSCLE Net) allows for specialized optimization. The MUSCLE Net's ability to focus on small candidate lesions is crucial for filtering artifacts.
- Patch-Based Processing for Imbalance: Using image patches as input helps address the significant class imbalance between healthy tissue and small lesions in DWI scans, which is a common challenge in medical image segmentation.
- DeconvNet's Advantage in Small Lesion Segmentation: The DeconvNet architecture, with its max pooling and unpooling strategy using pooling masks, is shown to be more effective than other CNN architectures (like FCN, DeepLab, U-Net) at reconstructing small lesions, which are often missed by methods relying on simpler upsampling techniques like bilinear interpolation.
- Importance of Patch Size: The study highlights that the size of the input image patches during training has a more significant impact on performance than the network's receptive field size. Medium-sized patches (64x64) provided the best balance.
Key Results
- Overall Performance: The proposed EDD + MUSCLE Net system achieved a mean Dice coefficient of 0.67 on the entire test dataset.
- Lesion Size Specific Performance:
- For subjects with small lesions, the mean Dice coefficient was 0.61.
- For subjects with large lesions, the mean Dice coefficient was 0.83.
- Detection Rate: The system achieved a high lesion detection rate (DR) of 0.94.
- False Positive Reduction: The MUSCLE Net significantly reduced false positives. For instance, on the validation dataset, the EDD Net produced an average of 8.87 false positives per image, which was reduced to 3.14 by the EDD + MUSCLE Net. The number of false positives was also significantly reduced in the small lesion subset (from 11.38 for the DeconvNet to 2.97 for EDD+MUSCLE Net).
- Comparison to Baselines: The EDD Net alone outperformed other baseline CNN architectures (DeepLab, FCN, U-Net) on the small lesion subset, achieving a Dice score of 0.56 compared to the best baseline's 0.47. The EDD + MUSCLE Net further improved this to 0.61.
Limitations and Future Work
- Non-End-to-End Training: The MUSCLE Net is trained separately from the EDD Net, meaning the entire system is not end-to-end. This limits efficiency and could potentially be improved.
- Focus on False Positives: While effective at removing false positives, the MUSCLE Net does not explicitly address false negatives. Future work could aim to correct these missed lesions.
- 2D vs. 3D Analysis: The current approach uses 2D slices due to anisotropy and potential interpolation errors in 3D resampling. Future work could explore 3D convolutions, possibly by incorporating image super-resolution techniques to handle anisotropic data.
- Dataset Size: While large, the authors suggest collecting more DW images for training and testing to further improve generalization.
- Artifact Handling: While the MUSCLE Net helps, DWI artifacts remain a challenge. Further research into more sophisticated artifact detection and removal could be beneficial.
Summary updated