A · REDNESS HEATMAPBACK · TODAY
PEAK 61MEAN 41AREA 4.2 cm²
HOT PIXELS 12%vs YESTERDAY −6
B · PATCH DETECTION
P1Upper-back · 2.8 cm²61▲ +4
P2Left shoulder · 1.4 cm²38▼ −9
Each patch scored on its own (Δ vs 7-day) so one spot doesn't muddy the other's trend.
WHAT THE MODEL MUST OUTPUT
1 · Skin / lesion mask
Per-pixel segmentation → feeds the heatmap (A), area cm² and patch boxes (B).
2 · Redness + texture per pixel
Calibrated a* (redness) & roughness, normalized with a color reference so lighting can't fake a flare.
3 · Embedding vector
A numeric fingerprint per photo → nearest-neighbour look-alikes (E) and "seen this before".
4 · Region + pose tag
Body part + aligned framing → clean time-lapse (C) and per-region trends.