Abstract
Quantitative trichoscopy requires reliable localization of individual hair shafts, but dense per-hair annotation is costly, especially for thin, overlapping structures. We propose a scale-aware synthetic compositing pipeline for hair-shaft instance segmentation. Reusable hair crops are placed onto inpainted hair-free scalp backgrounds, generating 20,000 synthetic images with exact per-hair instance masks. A follicle-diameter calibration factor normalizes scale during synthesis and inference across devices and magnification settings. Evaluated on 365 real trichoscopy images, a model trained exclusively on synthetic data achieves Box mAP50 = 0.52 and Jaccard = 0.54. A union of complementary inference variants improves Box mAP50 while matching the best single-variant Jaccard. Compared with a semantic segmentation baseline, our method reduces count MAE by 43% from 24.2 to 13.7 hairs and improves counting correlation from 0.87 to 0.93. These results suggest that scale-aware synthetic compositing can reduce annotation cost while enabling effective per-shaft localization.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/DGM4MICCAI_022.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/profile?id=~Mohamed_Ali_Ben_Youssef1
BibTex
@InProceedings{BenMoh_Synthetic_MICCAISAT2026,
author = { Ben Youssef, Mohamed Ali AND Dang, Khoi Tran AND Nguyen, Duc Thang AND Kelati, Awatef AND Nguyen, Hang},
title = { { Synthetic Data Generation for Automated Hair Instance Segmentation in Trichoscopy } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17258},
month = {pending},
page = {pending}
}