AI Follicle Classification & Matrix Generation
Hybrid YOLOv8-seg + U-Net pipeline classifies follicular units (FU1–FU4), estimates exit angle (±5°), and emits the 3D implantation matrix on the scalp depth surface.
A YOLOv8-seg backbone, trained on ≈ 1.4 M annotated scalp tiles across Fitzpatrick I–VI, produces per-follicle bounding masks at 60 ms inference (TensorRT FP16, RTX A4500). Each detection is routed through a lightweight 3-stage classifier head that assigns one of {FU1, FU2, FU3, FU4} based on shaft count and mutual spacing.
Exit angle and direction are estimated by a separate U-Net regression head that consumes the cross-polarized + IR channels and the local depth patch. Mean absolute angular error on the validation set is 4.7° (σ = 2.1°). Direction vectors are converted from camera frame to scalp-surface tangent frame using local LiDAR normals.
The 3D implantation matrix is built by Poisson-disk sampling on the scalp surface under hard constraints: minimum inter-implant spacing 0.8 mm, target density 25–40 FU/cm², and angular distribution matched to the native follicle field via von Mises–Fisher KDE.
| Parameter | Value | Unit | Tolerance / Note |
|---|---|---|---|
| Detection backbone | YOLOv8-seg | Custom 5-class head | |
| Classification head | 3-stage MLP | FU1/FU2/FU3/FU4 + reject | |
| Training corpus | ~1.4 M tiles | Fitzpatrick I–VI | |
| Inference latency | 60 | ms | TensorRT FP16 |
| [email protected] (detection) | 0.93 | Hold-out test set | |
| FU classification F1 | 0.91 | ||
| Angle MAE | 4.7 | ° | |
| Min inter-implant spacing | 0.8 | mm | |
| Target density | 25–40 | FU/cm² | |
| Sampler | Poisson-disk + vMF KDE |
Procedure-time disagreements between the YOLO head and the operator override are silently logged (PHI-stripped) and queued into the weekly retraining batch.