How arnoweng/chexnet works
Project type and purpose
• This is a machine-learning project (PyTorch-based) for classification and localization of thoracic diseases from chest X-ray images.
• Key purpose: train and evaluate a DenseNet-based CNN (CheXNet) on the NIH ChestX-ray14 dataset to output per-disease probabilities and heatmaps indicating pathology locations.
File structure
• Top level:
• ChestX-ray14/
– images/ (raw X-ray files)
– labels/ (train/val/test split lists)
• model.py (defines model, training, evaluation loops)
• read_data.py (data loading / preprocessing utilities)
• localization/ (generated likelihood maps / visualization outputs)
• model.pth.tar (pretrained or checkpoint weights)
• README.md (project description)
• Patterns: a single-service ML pipeline (data loader → model → results). No separate “frontend” or microservices.
README for additional insights
• Confirms dataset splits, model architecture (CheXNet / DenseNet121), training strategy (ten-crop augmentation), outputs (AUROC metrics, localization heatmaps).
Elements
• Clear labels for each component (e.g., “Data Loader”, “CheXNet Model”, “Localization Heatmap Generator”).
Detailed breakdown to capture essential architectural elements
1. Data Storage (ChestX-ray14/)
• Raw image files
2. Data Loading & Preprocessing
• read_data.py: parses labels, builds PyTorch Dataset
• DataLoader: batch loading, on-the-fly augmentation (e.g., ten-crop)
3. Model Component
• model.py: defines CheXNet architecture (DenseNet121 backbone)
• Training loop: loss computation, backprop, optimizer step
• Evaluation loop: compute AUROC per disease
4. Checkpoint Management
• Save best model to model.pth.tar
• Load pretrained weights if available
5. Localization Module
• After forward pass, generate likelihood map heatmaps
• Save PNGs under localization/
6. External Interfaces
• GPU / CUDA
• Python RT logs or console outputs for metrics
[ChestX-ray14/images + labels] → [Data Loader (read_data.py)] → [CheXNet Model (model.py) on GPU] → split outputs →
• [Metrics Calculator (AUROC)]
• [Localization Heatmap Generator → localization/]
↕
[Checkpoint Store (model.pth.tar)]