Features

A complete ML pipeline for medical wound segmentation, from training to production deployment.

U-Net Training

Train 3-level U-Net encoder-decoder models on custom wound datasets with skip connections for precise boundary detection.

Transfer Learning

Fine-tune pre-trained models on new datasets. Start from baseline checkpoints and adapt to institution-specific data.

Threshold Calibration

Grid-search optimization to find optimal operating thresholds that maximize Dice coefficient on test sets.

3-State Classification

Clinical decision support with WOUND, NO_WOUND, and UNCERTAIN states. Borderline cases flagged for human review.

ONNX Export

Convert trained PyTorch models to ONNX format with bundled metadata for cross-platform production deployment.

Batch Inference

Process single images or entire directories with optional probability heatmap generation for analysis.

Gallery Visualization

Generate HTML/image grids showing TP/FN/FP/TN examples with mask overlays for performance analysis.

Policy Locking

Persist decision parameters for reproducible deployments across environments. Governance and audit support.

Secure Model Loading

Safe model loading with weights_only=True prevents arbitrary code execution. Path traversal protection included.

Typical Workflow

1

Train

Train U-Net model on your dataset

2

Calibrate

Find optimal threshold via grid search

3

Analyze

Lock policy for reproducibility

4

Deploy

Export to ONNX for production