🧠 RSNA Intracranial Aneurysm Detection

A deep learning project built for the RSNA Intracranial Aneurysm Detection competition on Kaggle.
This repository explores 3D medical image classification for detecting intracranial aneurysms in CT angiography (CTA) scans, using a pipeline designed for accuracy, efficiency, and reproducibility.

---

🚀 Project Overview

The goal of this project is to automatically identify aneurysms within 3D DICOM series of the brain.
Each sample represents a full CTA volume, and the model must determine whether an aneurysm is present.

This is a highly imbalanced classification problem — only a small subset of patients have positive findings — which makes data sampling and augmentation key to achieving strong results.

---

🧩 Dataset

• Source: RSNA Intracranial Aneurysm Detection
• Data format: DICOM (3D CT angiography)
• Accompanying CSVs:
  • train.csv: patient-level aneurysm labels
  • train_localizers.csv: slice indices and localization metadata
  • Segmentation masks for annotated aneurysms

The dataset is large, and pre-processing is crucial.
I used optimized DICOM parsing, volume normalization, and windowing (brain window) to extract relevant signal intensity for model input.

---

⚙️ Technical Stack

• Language: Python 3.10
• Framework: PyTorch (with MONAI transforms)
• Environment: Kaggle Notebook & Local GPU
• Tools: NumPy, pandas, nibabel, tqdm, scikit-learn, Matplotlib
• Augmentation: Random flips, rotations, intensity normalization
• Hardware: A100 / RTX local GPU

---

🧠 Model Design

This project evolved from a baseline 3D CNN to a hybrid architecture optimized for both speed and accuracy.

Versions Explored:

1. 3D ResNet18 baseline

   • Trained on preprocessed 3D patches
   • BCEWithLogits loss + class weighting

2. Stratified Oversampling + Focal Loss

   • Increased exposure to rare positive cases
   • Improved sensitivity without overwhelming the model

3. Patch-based Ensemble (WIP)

   • Aggregates predictions from multiple 3D crops per scan
   • Aims to boost stability and AUC

---

📈 Results & Metrics

Experiment	Model	Sampling	AUC	Notes
v1	3D ResNet18	None	0.74	Baseline
v2	3D ResNet18	Stratified oversampling	0.79	Reduced class imbalance
v3	3D ResNet34 + Focal Loss	Yes	0.82	Better sensitivity on rare positives

Metrics are computed using validation folds split by patient ID to prevent data leakage.

---

🧪 Training Pipeline

train.py manages the full training lifecycle:

• DICOM loading → 3D volume assembly
• Windowing and normalization
• Augmentations
• Batch generation with custom Dataset and DataLoader
• Multi-GPU training support
• Fold-based validation and AUC tracking

python train.py --epochs 20 --batch-size 8 --fold 0 --model resnet3d34