How rvn5801/skin-lesion-segmentation-with-deep-supervision-and-attention-mechanisms works
Project type and purpose
• This is a specialized deep-learning pipeline (an open-source tool) for skin-lesion segmentation using a DS-AttentionUNet model.
• It provides end-to-end functionality: data loading/preprocessing, model definition, training, batch inference, single-image inference, metric computation, and visualization.
File structure
• Top-level scripts/modules:
• config.py – global paths, hyperparameters
• dataset.py – data loading, augmentation
• unet.py – network architecture (DS-AttentionUNet)
• metrics.py – loss functions and evaluation metrics
• visualization.py – plotting utilities, overlay generation
• train.py – standalone training script
• predict.py – batch inference script
• inference.py – single-image inference script
• main.py – unified CLI entrypoint for train/predict/inference
• *.py (others) – helpers (dataset, metrics, visualization)
• No separate “frontend/backend”; it is a monolithic Python package with clearly delineated modules.
• External dependencies in requirements.txt (PyTorch, torchvision, albumentations, OpenCV, etc.).
• Output directory structure under “outputs/” for saved models, logs, visualizations, and final results.
README for additional insights
• Architecture: DS-AttentionUNet with attention gates, deep supervision, group normalization.
• Training strategies: Mixed precision, OneCycleLR, composite loss.
• Workflows:
• Training → outputs/models, logs, visualizations
• Batch prediction → results/ directory
• Single inference → results/ directory
a. Main components
1. Data Storage (ISIC 2018 dataset on disk)
2. Configuration Loader (config.py)
3. Dataset Module (dataset.py)
4. Model Module (unet.py)
5. Training Engine (train.py & train logic in main.py)
6. Inference Engine (predict.py & inference.py called via main.py)
7. Metric & Loss Module (metrics.py)
8. Visualization Module (visualization.py)
9. CLI Orchestrator (main.py)
10. External Libraries & Hardware (PyTorch on GPU)
11. Output Storage (outputs/… and results/)
b. Relationships & interactions
• CLI Orchestrator parses commands → loads Config → invokes Dataset Module and Model Module.
• For training: Dataset → Training Engine → Model → Metric Module → Visualization → Output Storage.
• For inference: Input images → Inference Engine → Model → Visualization → results/.
• All modules rely on external libraries and hardware (GPU, CUDA).
c. Architectural patterns & design principles
• Modular separation of concerns: data, model, metrics, visualization, entrypoint.
• Reusable components: dataset and model modules are shared by train/predict/inference.
• Single CLI interface (main.py) for consistency and extensibility.
d. Technologies & frameworks
• PyTorch (+ torchvision), Albumentations for augmentation, OpenCV for image I/O, Matplotlib for plots, NumPy/pandas for data handling.
Key points to capture
• The unified CLI (main.py) at the top, branching to train/predict/inference workflows.
• Shared core modules (dataset.py, unet.py, metrics.py, visualization.py).
• Data flow and artifact generation (models, logs, visualizations, results).
• External integration points (dataset on disk, GPU acceleration).
• Ability to extend: drop-in new architectures, augmentations, or loss functions by modifying respective modules.