Upload CHANGELOG.md with huggingface_hub

CHANGELOG.md

@@ -0,0 +1,106 @@
+# Changelog
+
+All notable changes to DiaFoot.AI are documented in this file.
+
+## [2.0.0] — 2026-03-05
+
+### Complete Rebuild: v1 → v2
+
+DiaFoot.AI v2 is a ground-up rebuild addressing fundamental flaws in v1's data pipeline and evaluation methodology.
+
+### Why the Rebuild
+
+v1 achieved 84.93% IoU / 91.73% Dice but had zero clinical specificity — it predicted ulcers on every input because it was trained exclusively on ulcer images with no concept of "not a wound." v2 transforms the project from a toy segmentation model into a clinically meaningful multi-task pipeline.
+
+### Added
+
+**Data Foundation**
+- Three-category dataset: 2,119 DFU + 3,300 healthy + 2,686 non-DFU = 8,105 total images
+- Production data cleaning pipeline: integrity checks, mask validation, perceptual hash deduplication
+- CleanVision quality audit on all raw data (blur, darkness, duplicates, low-information detection)
+- ITA skin tone analysis for fairness evaluation
+- Stratified train/val/test splits (70/15/15) with zero leakage verification
+- AZH wound care center dataset integration (1,109 new wound images)
+
+**Architecture**
+- Multi-task cascaded pipeline: Triage Classifier → Wound Segmenter
+- EfficientNet-V2-M triage classifier (healthy vs non-DFU vs DFU)
+- U-Net++ with EfficientNet-B4 encoder and scSE attention
+- FUSegNet with EfficientNet-B7 and P-scSE attention (comparison architecture)
+- MedSAM2 LoRA fine-tuning setup (implemented, not trained)
+- nnU-Net v2 wrapper (implemented, not trained)
+
+**Training**
+- DiceCE compound loss function
+- Cosine annealing scheduler with linear warmup
+- Exponential Moving Average (EMA) weight tracking
+- Early stopping (patience=15 epochs)
+- BFloat16 mixed precision on H200 GPUs
+- SLURM array jobs for parallel ablation studies
+
+**Evaluation**
+- Full metrics suite: Dice, IoU, HD95, NSD@2mm, NSD@5mm, ASSD
+- Clinical metrics: wound area estimation (mm²), wound perimeter
+- Test-Time Augmentation (TTA) with 16 augmentations (+3.88% Dice improvement)
+- ITA-stratified fairness audit (0.00% gap on DFU images)
+- Data composition ablation (DFU-only vs mixed training)
+- Architecture ablation (U-Net++ vs FUSegNet)
+
+**Deployment**
+- End-to-end inference pipeline (classify → segment → measure)
+- FastAPI REST API with /predict endpoint
+- ONNX export pipeline with validation and benchmarking
+- Prediction visualization with mask overlays
+
+**Documentation**
+- Comprehensive README with results tables and honest limitations
+- Dataset card documenting all sources and ITA distribution
+- 38-commit structured development plan mapped to peer feedback
+- Results notebooks with publication-ready figures
+
+### Changed
+
+- Training data: ulcer-only → three categories (healthy + non-DFU + DFU)
+- Architecture: single binary segmenter → cascaded multi-task pipeline
+- Evaluation: Dice/IoU only → Dice, IoU, HD95, NSD, clinical metrics, fairness
+- Loss function: Focal Tversky → DiceCE compound loss
+- Encoder: EfficientNet-B4 → EfficientNet-B4 with scSE attention
+- Training loop: basic loop → production trainer with EMA, early stopping, checkpointing
+- Data pipeline: raw images → cleaned, validated, deduplicated, stratified
+
+### Fixed
+
+- Model predicting ulcers on healthy skin (added negative examples)
+- No data quality assurance (added CleanVision + Cleanlab pipeline)
+- Inflated metrics from training on uncleaned data
+- No skin tone fairness analysis (added ITA-stratified evaluation)
+- No boundary quality metrics (added HD95, NSD)
+
+### Key Results
+
+| Metric | v1 | v2 | Improvement |
+|--------|----|----|-------------|
+| Dice (overall) | 91.73%* | 85.89% | Honest measurement |
+| IoU (overall) | 84.93%* | 79.35% | Honest measurement |
+| Clinical specificity | 0% | 100% (classifier) | From useless to useful |
+| HD95 | Not measured | 17.3 px | New metric |
+| NSD@5mm | Not measured | 94.74% | New metric |
+| Wound area error | Not measured | 1.1% | New metric |
+| Skin tone fairness gap | Not measured | 0.00% | New metric |
+| Data categories | 1 (DFU only) | 3 (DFU + healthy + non-DFU) | Clinically complete |
+| Data cleaning | None | Full pipeline | Production grade |
+
+*\*v1 metrics were inflated by training on uncleaned data without negative examples.*
+
+---
+
+## [1.0.0] — 2026-02-01
+
+### Initial Release
+
+- U-Net++ with EfficientNet-B4 encoder
+- FUSeg dataset (1,210 images, ulcers only)
+- Focal Tversky loss
+- 84.93% IoU, 91.73% Dice
+- Basic CLAHE preprocessing
+- No data cleaning, no negative examples, no fairness analysis