RSNA Intracranial Aneurysm Challenge 2025

1. Understanding the Challenge

The RSNA Intracranial Aneurysm Challenge is a multi-label classification task, where we aim to predict the presence of aneurysms across multiple anatomical regions in the brain. The competition also provides segmentation data, which we plan to leverage for backbone pretraining and fine-tuning.

Key Challenges:

• Heterogeneous data: Images come from different institutions with varying acquisition protocols.
• Multiple modalities: Includes CTA, MRI (T1, T2), and MRA data. However, each patient has only one modality, leading to class imbalance across modalities.
• Large-scale dataset: ~314 GB of DICOM data requiring efficient handling and preprocessing.
• Longitudinal/Sequential nature: The data represents volumetric scans (3D), so we must respect slice order (z-index) for learning anatomical context.

---

2. Research & Exploration

Our initial research focused on understanding medical imaging pipelines and 3D visualization. Topics explored include:

• DICOM tags such as ImagePositionPatient and SliceThickness for correct stacking.
• Visualization using 3D Slicer to explore voxel-based data.
• Use of MONAI for medical imaging preprocessing.
• Sequential modeling strategies for volumetric data.

📂 Research Materials: Found in research_materials/ folder.

---

3. Dataset

• Size: ~314 GB
• Format: DICOM series per patient
• Processing: Converted DICOM series to .npy arrays sorted by z-index.
• Normalization: Per-scan min-max normalization.
• Resizing: Resized slices to 224×224.
• Storage: Compressed embeddings occupy ~1.9 GB, enabling efficient loading.

---

4. Approach 1 – RadImageNet + GRU

Our first approach used RadImageNet pretrained embeddings + a Bidirectional LSTM/GRU recurrent model for classification.

Training Details:

• Hardware: NVIDIA T4 GPU (Colab)

• Epochs: 35

• Batch Size: 8

• Loss Function: Weighted Categorical Cross-Entropy

  • Anatomical region labels: weight = 1.0
  • Final aneurysm label: weight = 13.0

• Optimizer: Adam with weight decay 1e-5

• Learning Rate: 1e-4 with decay schedule

Metrics & Results:

• Best Validation AUC: 0.5880
• Final Test Score: 0.59

Some labels achieved high AUC, while others were poor (see per-label AUC table below). This indicates that our backbone may not be sufficiently specialized for aneurysm features.

RSNA Intracranial Aneurysm Challenge 2025 – Project Readme

1. Understanding the Challenge

The RSNA Intracranial Aneurysm Challenge is a multi-label classification task, where we aim to predict the presence of aneurysms across multiple anatomical regions in the brain. The competition also provides segmentation data, which we plan to leverage for backbone pretraining and fine-tuning.

Key Challenges:

• Heterogeneous data: Images come from different institutions with varying acquisition protocols.
• Multiple modalities: Includes CTA, MRI (T1, T2), and MRA data. However, each patient has only one modality, leading to class imbalance across modalities.
• Large-scale dataset: ~314 GB of DICOM data requiring efficient handling and preprocessing.
• Longitudinal/Sequential nature: The data represents volumetric scans (3D), so we must respect slice order (z-index) for learning anatomical context.

---

2. Research & Exploration

Our initial research focused on understanding medical imaging pipelines and 3D visualization. Topics explored include:

• DICOM tags such as ImagePositionPatient and SliceThickness for correct stacking.
• Visualization using 3D Slicer to explore voxel-based data.
• Use of MONAI for medical imaging preprocessing.
• Sequential modeling strategies for volumetric data.

📂 Research Materials: Found in research_materials/ folder.

---

3. Dataset

• Size: ~314 GB
• Format: DICOM series per patient
• Processing: Converted DICOM series to .npy arrays sorted by z-index.
• Normalization: Per-scan min-max normalization.
• Resizing: Resized slices to 224×224.
• Storage: Compressed embeddings occupy ~1.9 GB, enabling efficient loading.

---

4. Approach 1 – RadImageNet + GRU

Our first approach used RadImageNet pretrained embeddings + a Bidirectional LSTM/GRU recurrent model for classification.

Training Details:

• Hardware: NVIDIA T4 GPU (Colab)

• Epochs: 35

• Batch Size: 8

• Loss Function: Weighted Categorical Cross-Entropy

  • Anatomical region labels: weight = 1.0
  • Final aneurysm label: weight = 13.0

• Optimizer: Adam with weight decay 1e-5

• Learning Rate: 1e-4 with decay schedule

Metrics & Results:

• Best Validation AUC: 0.5880
• Final Test Score: 0.59

Some labels achieved high AUC, while others were poor (see per-label AUC table below). This indicates that our backbone may not be sufficiently specialized for aneurysm features.

Training History (Loss, AUC- More can be found in src\model_traininig\models\radImagenet+gru_0.59\artifacts)

---

5. Next Steps

• Improve Backbone: Move towards UNet-style architectures to leverage segmentation data for pretraining.
• Add Transformers: Incorporate vision transformers for better modeling of long-range dependencies.
• Better Augmentation: Explore modality-specific augmentation strategies.
• Hyperparameter Tuning: Optimize learning rate schedules, class weights, and regularization.

---

6. Artifacts

Model artifacts and logs are stored here:

src/model_traininig/models/radImagenet+gru_0.59

---

7. Key Takeaways

• The GRU approach demonstrated the importance of sequential context but plateaued at ~0.59 AUC.
• Backbone selection is critical — segmentation-aware models may boost performance.
• The multimodal and class-imbalanced nature of the data remains the primary challenge.

---

8. Future Work

• Implement multi-task learning (classification + segmentation).
• Experiment with 3D CNNs or Swin Transformers to better model volumetric data.
• Investigate domain adaptation techniques to handle inter-institutional variability.

---

5. Next Steps

• Improve Backbone: Move towards UNet-style architectures to leverage segmentation data for pretraining.
• Add Transformers: Incorporate vision transformers for better modeling of long-range dependencies.
• Better Augmentation: Explore modality-specific augmentation strategies.
• Hyperparameter Tuning: Optimize learning rate schedules, class weights, and regularization.

---

6. Artifacts

Model artifacts and logs are stored here:

src/model_traininig/models/radImagenet+gru_0.59

---

7. Key Takeaways

• The GRU approach demonstrated the importance of sequential context but plateaued at ~0.59 AUC.
• Backbone selection is critical — segmentation-aware models may boost performance.
• The multimodal and class-imbalanced nature of the data remains the primary challenge.

---

8. Future Work

• Implement multi-task learning (classification + segmentation).
• Experiment with 3D CNNs or Swin Transformers to better model volumetric data.
• Investigate domain adaptation techniques to handle inter-institutional variability.