The task involves detecting and localizing intracranial aneurysms from multimodal 3D medical imaging data (CTA / MRA / MRI), which is challenging due to:
\nMy approach focuses on improving signal-to-noise ratio through a localization-first 3D pipeline combined with multimodal modeling and ensembling.
\nLocalization-first 3D pipeline to reduce background noise and improve detection of small lesions.
\nInstead of directly performing classification on full-head scans, I first localize relevant anatomical regions and then apply downstream models on refined Regions of Interest (ROI).
\nThe solution is built as a three-stage pipeline:
\nTwo complementary model families are used:
\nTransformer-based 3D models
\nCNN + RNN + Attention models
\nPredictions from multiple models are combined using weighted ensembling, which improves robustness and stabilizes performance across validation folds.
\nThe solution shows strong performance on both:
\nFor full details, code, and system design:
\nGitHub Repository:
\nhttps://github.com/WoshengDeng/rsna-intracranial-aneurysm-detection
This writeup is intentionally concise. The full pipeline and implementation details are available in the GitHub repository.
", "messages": [ { "id": 3429701, "postDate": "2026-03-27T05:28:14.280Z", "content": "The task involves detecting and localizing intracranial aneurysms from multimodal 3D medical imaging data (CTA / MRA / MRI), which is challenging due to:
\nMy approach focuses on improving signal-to-noise ratio through a localization-first 3D pipeline combined with multimodal modeling and ensembling.
\nLocalization-first 3D pipeline to reduce background noise and improve detection of small lesions.
\nInstead of directly performing classification on full-head scans, I first localize relevant anatomical regions and then apply downstream models on refined Regions of Interest (ROI).
\nThe solution is built as a three-stage pipeline:
\nTwo complementary model families are used:
\nTransformer-based 3D models
\nCNN + RNN + Attention models
\nPredictions from multiple models are combined using weighted ensembling, which improves robustness and stabilizes performance across validation folds.
\nThe solution shows strong performance on both:
\nFor full details, code, and system design:
\nGitHub Repository:
\nhttps://github.com/WoshengDeng/rsna-intracranial-aneurysm-detection
This writeup is intentionally concise. The full pipeline and implementation details are available in the GitHub repository.
", "rawMarkdown": "The task involves detecting and localizing intracranial aneurysms from **multimodal 3D medical imaging data (CTA / MRA / MRI)**, which is challenging due to:\n\n- Small lesion size \n- Severe class imbalance \n- High variability across imaging modalities \n\nMy approach focuses on improving signal-to-noise ratio through a **localization-first 3D pipeline** combined with multimodal modeling and ensembling.\n\n---\n\n### Key Idea\n\n**Localization-first 3D pipeline to reduce background noise and improve detection of small lesions.**\n\nInstead of directly performing classification on full-head scans, I first localize relevant anatomical regions and then apply downstream models on refined Regions of Interest (ROI).\n\n---\n\n### Approach Overview\n\nThe solution is built as a **three-stage pipeline**:\n\n#### Stage 1 — Brain ROI Localization\n- A 3D segmentation model (MedNeXt-S) is applied to extract the brain region from full-head scans. \n- This removes irrelevant background and improves downstream efficiency. \n\n#### Stage 2 — Artery ROI Refinement\n- Within the brain region, a second-stage model focuses on arterial structures. \n- Anatomical priors (e.g., TotalSeg) are used to guide localization. \n\n#### Stage 3 — Aneurysm Classification\n\nTwo complementary model families are used:\n\n- Transformer-based 3D models \n - Operate on ROI inputs \n - Capture global spatial features \n\n- CNN + RNN + Attention models \n - Operate on volumetric sequences \n - Capture temporal/contextual patterns \n\n---\n\n### Ensemble Strategy\n\nPredictions from multiple models are combined using **weighted ensembling**, which improves robustness and stabilizes performance across validation folds.\n\n---\n\n### Result\n\n- Final Score: **0.70 Mean Weighted AUC-ROC** \n- Rank: **57 / 1147 (Top 5%)** \n\nThe solution shows strong performance on both:\n- Primary label (Aneurysm Present) \n- Auxiliary anatomical labels \n\n---\n\n### Full Implementation\n\nFor full details, code, and system design:\n\nGitHub Repository: \nhttps://github.com/WoshengDeng/rsna-intracranial-aneurysm-detection\n\n---\n\nThis writeup is intentionally concise. The full pipeline and implementation details are available in the GitHub repository." } ], "comments": [], "raw_markdown_by_id": { "3429701": "The task involves detecting and localizing intracranial aneurysms from **multimodal 3D medical imaging data (CTA / MRA / MRI)**, which is challenging due to:\n\n- Small lesion size \n- Severe class imbalance \n- High variability across imaging modalities \n\nMy approach focuses on improving signal-to-noise ratio through a **localization-first 3D pipeline** combined with multimodal modeling and ensembling.\n\n---\n\n### Key Idea\n\n**Localization-first 3D pipeline to reduce background noise and improve detection of small lesions.**\n\nInstead of directly performing classification on full-head scans, I first localize relevant anatomical regions and then apply downstream models on refined Regions of Interest (ROI).\n\n---\n\n### Approach Overview\n\nThe solution is built as a **three-stage pipeline**:\n\n#### Stage 1 — Brain ROI Localization\n- A 3D segmentation model (MedNeXt-S) is applied to extract the brain region from full-head scans. \n- This removes irrelevant background and improves downstream efficiency. \n\n#### Stage 2 — Artery ROI Refinement\n- Within the brain region, a second-stage model focuses on arterial structures. \n- Anatomical priors (e.g., TotalSeg) are used to guide localization. \n\n#### Stage 3 — Aneurysm Classification\n\nTwo complementary model families are used:\n\n- Transformer-based 3D models \n - Operate on ROI inputs \n - Capture global spatial features \n\n- CNN + RNN + Attention models \n - Operate on volumetric sequences \n - Capture temporal/contextual patterns \n\n---\n\n### Ensemble Strategy\n\nPredictions from multiple models are combined using **weighted ensembling**, which improves robustness and stabilizes performance across validation folds.\n\n---\n\n### Result\n\n- Final Score: **0.70 Mean Weighted AUC-ROC** \n- Rank: **57 / 1147 (Top 5%)** \n\nThe solution shows strong performance on both:\n- Primary label (Aneurysm Present) \n- Auxiliary anatomical labels \n\n---\n\n### Full Implementation\n\nFor full details, code, and system design:\n\nGitHub Repository: \nhttps://github.com/WoshengDeng/rsna-intracranial-aneurysm-detection\n\n---\n\nThis writeup is intentionally concise. The full pipeline and implementation details are available in the GitHub repository." } }