AlgomCorp/hair-segmentation-unet

Hair Progress Tracker & Deep Learning Pipeline Visualizer

Deep Learning Capstone Project | Clinical Scalp Hair Segmentation & Progression Tracking

---

🌟 Overview & Clinical Utility

The Hair Progress Tracker is a clinical-grade deep learning system designed for quantitative, objective monitoring of scalp hair progression (alopecia, hair loss treatments, hair transplant recovery, and regrowth therapy).

Core Engineering Capabilities (v2.0):

1. Accurate U-Net Semantic Segmentation:
   • Symmetrical U-Net with channels [32, 64, 128, 256] and bottleneck 512 (~7.76M parameters).
   • Benchmarked on 2,930 holdout test images: Test Dice: 95.19%, Test IoU: 90.84%, Precision: 94.88%, Recall: 95.51%.
2. Test-Time Augmentation (TTA):
   • Dual-pass mirror-flip probability averaging eliminates boundary edge jitter and stabilizes hairline segmentation.
3. Hierarchical Multi-Tier Registration:
   • Tier 1 (Facial Biometric Alignment): Rigid similarity transform mapping facial centers and pupil geometry.
   • Tier 2 (AKAZE + RANSAC): Feature matching with affine determinant sanity checks ($0.4 \le \det(H) \le 2.5$) for crown/vertex views.
   • Tier 3 (Reference Frame): Evaluates unaligned photos with true positive structural similarity ($0.0 \le \text{SSIM} \le 1.0$).
4. Scale-Invariant Cranium Normalization:
   • Computes hair density relative to the patient's anatomical head region ($\frac{N_{\text{hair}}}{N_{\text{head}}}$), guaranteeing invariance to camera zoom and distance shifts.
5. Continuous Scalp Follicle Density Heatmap:
   • Turbo colormap of continuous sigmoid probabilities identifies regions of sub-clinical thinning before complete follicular loss occurs.
6. Dual Hairline Shift Contours:
   • Simultaneous visualization of Baseline (Green) and Follow-up (Blue) hairlines with vertical pixel displacement tracking.
7. Clinical Trichology Staging:
   • Automatic classification on the Norwood-Hamilton Scale (Types I–VII) for male pattern hair loss.
8. Interactive Split Curtain Slider:
   • Scrubbable $0% \to 100%$ before/after wipe viewer.
9. Mathematical Provenance & Pixel Audit Inspector:
   • Live on-screen verification of raw pixel counts ($65,536\text{ px}$ frame grid), formulas, and tensor metadata proving 100% determinism.
10. Medical-Grade 6-Panel Report Export:
    • Generates a publication-grade summary report PNG.