---

Blindsweep GA Prediction

---

A clean and modular deep learning template to predict Gestational Age (GA) from Blindsweep ultrasound datasets using PyTorch. Built for reproducibility, simplicity, and extensibility. 🚀🧠

---

📌 Introduction

Why you might want to use it:

✅ Ready-to-use template Includes datasets, models, training loops, and inference pipelines for GA prediction.

✅ Attention-based video aggregation Uses a ResNet backbone and weighted average attention to handle multiple frames per sweep.

✅ Educational and reproducible Thoroughly commented code and modular dataset/model classes make it easy to learn and extend.

---

Why you might not want to use it:

❌ Not optimized for very large datasets (can be adapted).

❌ GPU is recommended; CPU training will be slow.

---

🛠 Features

• Handles single and multi-sweep datasets.
• Uses ResNet18 backbone with optional fine-tuning.
• Weighted average attention for frame aggregation.
• Training and validation with MAE loss.
• Saves best model automatically.
• TensorBoard integration for visualization.

---

⚙️ Environment Setup

1. Install Miniconda

wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
bash Miniconda3-latest-Linux-x86_64.sh
echo 'eval "$($HOME/miniconda3/bin/conda shell.bash hook)"' >> ~/.bashrc
source ~/.bashrc

2. Create & activate environment

conda create -n ga-us python=3.10
conda activate ga-us

3. Install dependencies

pip3 install --pre torch torchvision --index-url https://download.pytorch.org/whl/nightly/cu128
pip install notebook pandas tensorboard
conda install -c conda-forge nibabel
pip install tqdm

4. Create project directories

mkdir checkpoints logs

---

📂 Dataset

• Train dataset: single sweep per sample.
• Validation/Test dataset: multiple sweeps per sample.
• CSV files should include:

study_id, ga, path_nifti_1, path_nifti_2, ...

---

📂 Folder Structure

.
├── .gitignore
├── check.ipynb
├── data.py
├── evaluate_metrics.py
├── infer.py
├── model.py
├── requirements.txt
├── train.py
└── checkpoints/
    └── best_model.pth
└── logs/
    ├── events.out.tfevents...

• .gitignore: Git ignore file for unnecessary files.
• check.ipynb: Jupyter notebook for exploratory analysis or debugging.
• data.py: Dataset loading and preprocessing logic.
• evaluate_metrics.py: Code to evaluate model performance.
• infer.py: Inference logic for prediction.
• model.py: Contains model architecture (ResNet backbone).
• requirements.txt: List of required Python packages.
• train.py: Training loop and validation logic.
• checkpoints/: Directory to store saved model checkpoints.
• logs/: Directory to store TensorBoard logs.

---

🏗 Model Architecture

• Backbone: ResNet18 (pretrained, optionally frozen).
• Attention: WeightedAverageAttention for frame aggregation.
• Output: Linear layer predicting GA.

Improve the model as required.

---

🚀 Training & Validation

Run training using:

from train import train_and_validate

train_csv = "path/to/train.csv"
val_csv = "path/to/val.csv"
train_and_validate(train_csv, val_csv, epochs=100, batch_size=8, n_sweeps_val=8, save_path='checkpoints/best_model.pth')

• Uses MAE (L1) loss. The code can be adapted to use other loss functions.
• Saves best model automatically to checkpoints/best_model.pth.

TensorBoard Integration

You can visualize the training process with TensorBoard. To log training metrics and visualize them:

1. Start TensorBoard on the server:

   Run the following command on the server where your model is training. This will start the TensorBoard service:

   tensorboard --logdir=logs --port=6000

   • --port=6000 specifies the port to use (default is 6006).

2. Map the server port to your local machine:

   If you're connecting to the server remotely via SSH, you'll need to forward the TensorBoard port so you can access it locally. In your terminal (on your local machine), run:

   ssh -L -p 5555 6000:localhost:6000 user@server_ip

   Replace user@server_ip with your actual username and server IP. This command forwards the server's port 6006 (where TensorBoard is running) to your local machine's port 6006.

3. Open TensorBoard in your browser:

   Once the SSH tunnel is established, open your browser and navigate to http://localhost:6006 to view TensorBoard and monitor metrics like loss, accuracy, etc.

---

🧪 Inference

To predict GA from test data:

from infer import predict_ga

model_path = "checkpoints/best_model.pth"
test_csv = "path/to/test.csv"
predictions = predict_ga(model_path, test_csv)

Example Inference CSV

Here’s a sample of how your inference CSV should look:

study_id, site, predicted_ga
KA-PC-002-1, Kenya, 180
NL-PC-087-1, Nepal, 157
PN-PC-090-1, Pakiastan, 223

---

📖 References

• PyTorch
• TorchVision
• NiBabel for NIfTI image handling

---

📝 License

MIT License.

---