Ischemic Stroke Lesion Segmentation on ISLES 2022

A comparative study of 3D deep learning architectures for multimodal MRI stroke lesion segmentation, including a proposed separate-encoder U-Net that achieved the highest mean Dice score.

Dataset: ISLES 2022 — Multi-center MRI Stroke Lesion Segmentation  | 

---

🧠 Project Overview

Ischemic stroke is a leading cause of disability worldwide. Accurate segmentation of stroke lesions from MRI is critical for diagnosis and treatment planning. This project benchmarks four 3D deep learning architectures on the ISLES 2022 dataset and proposes a separate-encoder multimodal U-Net that processes DWI, ADC, and FLAIR independently before feature fusion — achieving the best mean Dice among all tested models.

📑 Technical Report

A full technical report documenting the complete methodology, results, and discussion is available here:

🔗 View Technical Report (PDF)

---

📂 Dataset

• ISLES 2022: 400 multi-center MRI cases (250 training, 150 test)
• Modalities: DWI, ADC, FLAIR (.nii.gz)
• Centers: Three stroke centers in Germany and Switzerland (Philips & Siemens, 1.5T and 3T)
• Ground truth: Manually verified lesion masks via ITK-SNAP and 3D Slicer

---

⚙️ Preprocessing

• Resampled all modalities to uniform 1×1×1 mm voxel spacing
• Resized volumes to fixed 128×128×128 spatial dimensions
• Intensity normalisation using 1st–99th percentile clipping
• FLAIR registered to DWI space; skull stripping via HD-BET
• Stacked input: DWI + ADC + FLAIR concatenated as a 3-channel volume (3×128×128×128)
• Separate input: each modality fed independently into its own encoder

---

🏗️ Architectures

1. 3D U-Net (Stacked Input)

Standard encoder-decoder with skip connections. All three modalities stacked as a single 3-channel volume input.

2. 3D Attention U-Net (Stacked Input)

Adds attention gates to the U-Net decoder to focus feature learning on lesion-relevant regions.

3. SegResNet (Stacked Input)

3D residual network from MONAI with residual blocks for more stable gradient flow and optimisation.

4. Separate-Encoder U-Net (Proposed)

Each modality (DWI, ADC, FLAIR) passes through its own independent encoder (4 blocks: 16→32→64→128 channels). All three encoder outputs are concatenated at a shared bottleneck (384→256 channels) and decoded through a single decoder with multi-encoder skip connections fused at each level. This preserves modality-specific features before fusion, unlike stacked models which mix modalities at layer 1.

---

📈 Results

The separate-encoder U-Net achieved the highest mean Dice (0.5628). SegResNet achieved the best single-fold score (0.5936). Results are consistent with the ISLES 2022 challenge benchmark range of 0.4–0.6 for single models.

Per-case Dice score distribution across 250 training cases (3-fold ensemble).

---

🔬 Qualitative Results

Subject 0161: Ground truth vs 3D U-Net multimodal vs 3D U-Net stacked predictions.

---

💡 Key Finding

Stacked-input models treat DWI, ADC, and FLAIR as interchangeable channels (like RGB), mixing them at the very first layer. However, these modalities carry fundamentally different medical information — DWI is highly sensitive to early ischemic injury, ADC provides quantitative diffusion information, and FLAIR highlights edema while suppressing CSF. The separate-encoder design allows each modality to develop independent representations before late fusion, leading to better segmentation performance.

---

🗂️ Repository Structure

ischemic-stroke-lesion-segmentation/
├── 3D_U-net_MIA.ipynb               ← 3D U-Net (stacked)
├── 3D_Unet_separate.ipynb           ← Separate-Encoder U-Net (proposed)
├── attentionUnet_ISLES2022.ipynb    ← Attention U-Net
├── SegResNet_ISLES2022.ipynb        ← SegResNet
├── assets/
│   ├── subject0161_comparison.png
│   ├── results_table.png
│   └── dice_distribution.png
└── README.md

---

📄 References

1. de la Rosa, E., Reyes, M., Liew, S. L., Hutton, A., Wiest, R., Kaesmacher, J., ... & Robben, D. (2025). DeepISLES: a clinically validated ischemic stroke segmentation model from the ISLES'22 challenge. Nature Communications. https://www.nature.com/articles/s41467-025-62373-x
2. Rahman, A., Chowdhury, M. E. H., & Wadud, M. S. I. (2025). Deep learning-driven segmentation of ischemic stroke lesions using multi-channel MRI. Biomedical Signal Processing and Control, 105, 107676. https://www.sciencedirect.com/science/article/pii/S1746809425001879
3. Hernandez Petzsche, M. R., de la Rosa, E., Hanning, U., Wiest, R., Valenzuela, W., Reyes, M., ... & Robben, D. (2022). ISLES 2022: A multi-center magnetic resonance imaging stroke lesion segmentation dataset. arXiv preprint. https://arxiv.org/abs/2209.09546
4. Hernandez Petzsche, M. R., de la Rosa, E., Hanning, U., Wiest, R., Valenzuela, W., Reyes, M., ... & Robben, D. (2022). ISLES 2022: A multi-center magnetic resonance imaging stroke lesion segmentation dataset. Scientific Data, Nature. https://www.nature.com/articles/s41597-022-01875-5

---

🛠️ Tools & Technologies

• Python, PyTorch, MONAI
• SimpleITK, NiBabel
• ITK-SNAP, 3D Slicer (visualisation)
• 3-fold cross-validation, 30 epochs per fold
• Dice loss, learning rates: 5×10⁻⁴ (U-Net variants), 1×10⁻⁴ (SegResNet)