Abstract
Acute infarcts are more accurately segmented using ensembles of convolutional neural networks trained with multiparametric maps than by using a single model trained with a solo map. Automated lesion segmentation has high agreement with manual techniques for identifying patients with large lesion volumes.
Introduction
In the management of acute ischemic stroke, "time is brain." Rapid and accurate measurement of infarct volume—the area of dead or dying brain tissue—is critical for determining treatment eligibility, such as mechanical thrombectomy. While manual segmentation by experts remains the gold standard, it is time-consuming and subjective, making it impractical for rapid emergency triage.
Existing automated methods often rely solely on Diffusion-Weighted Imaging (DWI). However, single-parameter maps can be noisy, potentially misidentifying artifacts or chronic lesions as acute strokes. This study investigates whether combining multiple MRI parameters (multiparametric) and multiple models (ensembles) can achieve human-level accuracy in lesion segmentation.
Methods
The researchers employed DeepMedic, a 3D convolutional neural network (CNN) architecture designed for medical imaging. DeepMedic utilizes two parallel pathways—one at normal resolution and one at lower resolution—to capture both fine local details and broader anatomical context.
Key innovations in this study include:
- Multiparametric Input: Models were trained on combinations of DWI, Apparent Diffusion Coefficient (ADC) maps, and low b-value (b=0) images (LOWB).
- Ensemble Learning: Instead of a single model, the researchers trained five independent CNNs and averaged their voxel-wise predictions. This "wisdom of the crowd" approach reduces false positives and individual model errors.
- Clinical Validation: The model's performance was evaluated on its ability to categorize patients into volume groups (e.g., >31 cm³ or >70 cm³) used in clinical trials like DAWN and DEFUSE 3.
The study utilized a training cohort of 116 subjects and an evaluation cohort of 151 subjects. The evaluation cohort was particularly challenging, spanning 16 years of varying MRI hardware and protocols, providing a rigorous test of the model's generalizability.
Results
The ensemble approach significantly outperformed single-parameter models. The best-performing ensemble (E3), which utilized all three MRI parameters (DWI, ADC, and LOWB), achieved a median Dice score of 82.2%.
Key findings include:
- High Correlation: Automated lesion volumes showed a strong correlation with manual expert volumes (Spearman ρ = 0.91).
- Clinical Utility: The agreement between AI and human experts for identifying large lesion volumes was excellent (Cohen’s κ = 0.86–0.90).
- Efficiency: The full ensemble pipeline completes segmentation in approximately 5 minutes, offering a significant speed advantage over manual tracing.
- Value of LOWB: Including low b-value images improved performance by helping the model identify early vasogenic edema, providing complementary data to DWI and ADC.
Conclusion
This study demonstrates that an ensemble of 3D-CNNs using multiparametric MRI data provides a robust and accurate tool for automated stroke lesion segmentation. While the model showed some limitations with very small lesions (<1 cm³), its high performance on larger, clinically significant lesions suggests strong potential for real-world triage. Future work may focus on improving detection for small lesions and enhancing the robustness of initial preprocessing steps like brain extraction.
Summary updated