This code provides the implementation of the model proposed in the RSNA Intracranial Aneurysm Detection Competition (submitted on Oct. 14, 2025), including some utility methods.

• Features
• How it Works
• Model Info
  • Training Stats
  • Architecture
  • Training Information
• Example Usage

Features

• Automatic model download from Hugging Face Hub.
• Flexible inference: Accepts both single and batched inputs, in NumPy or PyTorch tensor formats.
• Preprocessing included: Handles grayscale conversion and channel stacking internally.
• Output as DataFrame: Returns predictions as a pandas DataFrame with named columns for each anatomical region.

↑ Top

How it Works

The main class, AneurysmDetection, loads a pretrained 3D ResNet model and provides a unified method (predict_tensor) for inference. The method accepts:

• A single input with shape (C, D, H, W)
• A batch of inputs with shape (B, C, D, H, W)
• A list of inputs, each with shape (C, D, H, W)

The model expects volumetric data (e.g., MRI or CT) with:

• C: Number of channels (usually 1 for grayscale)
• D: Number of slices/frames (default: 16)
• H, W: Height and width (default: 224x224)

↑ Top

Model Info

Training Stats

Architecture

Component	Description
Backbone	VideoResNet (R3D-18)
├─ Stem	Conv3d(3→64, kernel=3×7×7, stride=(1,2,2)) → BN3d → ReLU
├─ Layer1	2 × BasicBlock(64→64) Each: Conv3d(3×3×3, stride=1) → BN3d → ReLU
├─ Layer2	2 × BasicBlock(64→128) • Block 1: stride=2 + downsample(64→128) • Block 2: stride=1
├─ Layer3	2 × BasicBlock(128→256) • Block 1: stride=2 + downsample(128→256) • Block 2: stride=1
├─ Layer4	2 × BasicBlock(256→512) • Block 1: stride=2 + downsample(256→512) • Block 2: stride=1
├─ AvgPool	AdaptiveAvgPool3d(output=1×1×1)
└─ FC	Identity (ResNet final FC removed)
Classifier	MLP: Linear(512→512) → BN1d → ReLU → Dropout(0.3) Linear(512→256) → BN1d → ReLU → Dropout(0.3) Linear(256→14)

Training Information

Model name	Total parameters	Target label	Best checkpoint	Validation Score	Aneurysm AUC	Mean AUC	Optimizer	Scheduler	Learning rate	Weight decay	Accumulation	Early stopping	Class weight	Combined loss	Cross-validation
r3d_18_aneurysm	33,565,390	Aneurysm Present	Epoch 14/30 ⭐	0.91781	0.99916	0.83645	AdamW	CosineAnnealingLR	1e-4	1e-4	4	5	3.0	Yes (focal=0.3)	10-fold (k=10)

↑ Top

Example Usage

To get started, clone the repository using the command below:

!git clone https://github.com/NeoGreenCode/AneurysmDetection.git

from AneurysmDetection import AneurysmDetection as A_model
import numpy as np

# Initialize the model (downloads weights automatically if not present)
a = A_model(model_name="R3D18_AneurysmDetection_33M")

# Create a dummy batch of 10 volumes, each with shape (C=1, D=16, H=224, W=224)
dummy_array = np.random.randn(2, 1, 16, 224, 224).astype(np.float32)

pred = a.predict_tensor(dummy_array)
pred

	Left Infraclinoid Internal Carotid Artery	Right Infraclinoid Internal Carotid Artery	Left Supraclinoid Internal Carotid Artery	Right Supraclinoid Internal Carotid Artery	Left Middle Cerebral Artery	Right Middle Cerebral Artery	Anterior Communicating Artery	Left Anterior Cerebral Artery	Right Anterior Cerebral Artery	Left Posterior Communicating Artery	Right Posterior Communicating Artery	Basilar Tip	Other Posterior Circulation	Aneurysm Present
0	0.052423	0.081153	0.247672	0.140270	0.137513	0.159553	0.133901	0.046614	0.042221	0.024711	0.041733	0.040193	0.046926	0.598635
1	0.052246	0.081089	0.251525	0.145575	0.136922	0.160558	0.136903	0.045311	0.042342	0.024428	0.042559	0.040697	0.047657	0.602177

↑ Top