philldevcoder/Hierarchical-CXR-Net

Training Details

Training Data

• Open-access chest X-ray datasets (e.g., NIH ChestX-ray14, CheXpert).
• Data preprocessing: normalization, resizing, augmentation.

Training Procedure

• Stage 1: EfficientNet-B0 for coarse classification (normal vs abnormal).
• Stage 2: EfficientNet-B2 for expert-level multi-label disease classification.
• Grad-CAM integrated for visual interpretability.

Training Hyperparameters

• Mixed precision (fp16)
• Optimizer: AdamW
• Learning rate scheduler: CosineAnnealing
• Loss: Weighted BCE with logits

---

Evaluation

Testing Data

• Evaluated on public benchmark datasets (CheXpert, NIH ChestX-ray14).

Metrics

• AUROC (per-class and mean)
• F1-score
• Sensitivity/Specificity

Results

• Mean AUROC ≈ 0.85–0.90 (depending on dataset and task)
• Grad-CAM heatmaps align with radiologically relevant regions

Model Examination

• Grad-CAM visualizations available for each prediction
• Two-stage pipeline mirrors clinical workflow

---

Environmental Impact

• Hardware Type: NVIDIA Tesla V100 (cloud GPU)
• Hours used: ~60 GPU hours
• Cloud Provider: Google Cloud
• Compute Region: US-central
• Carbon Emitted: Estimated ~25 kg CO2eq

---

Technical Specifications

Model Architecture

• Stage 1: EfficientNet-B0
• Stage 2: EfficientNet-B2
• Hierarchical classification pipeline
• Grad-CAM interpretability module

Compute Infrastructure

• Hardware: NVIDIA V100 (16GB)
• Software: PyTorch, Hugging Face Transformers, CUDA 11.8

---

Citation

BibTeX

@article{indabax2025cxrnet,
  title={Hierarchical CXR-Net: A Two-Stage Framework for Efficient and Interpretable Chest X-Ray Diagnosis},
  author={Ssempeebwa, Phillip and IndabaX Uganda AI Research Lab},
  year={2025},
  journal={Digital Health Africa 2025 Poster Proceedings}
}