A General Deep Learning Framework for Automatic Multi-View Facial and Nasal Landmark Detection in Clinical Photographs

Abstract

Facial and nasal anatomical landmarks are essential for quan-
titative analysis and planning in aesthetic rhinoplasty, yet
manual annotation of multi-view clinical photographs is time-
consuming and subject to inter-observer variability. We
present a unified deep learning framework for automatic land-
mark detection across frontal, lateral, and basal views us-
ing a clinically motivated taxonomy of 42 landmarks. The
taxonomy explicitly distinguishes bilateral points and re-
solves basal-view alare-prime into superior and inferior vari-
ants to reduce geometric ambiguity. The dataset contains
1,217 clinical facial images from approximately 400 sub-
jects, with subject-wise splits into training/validation/test sets
(972/124/121). Landmark detection is formulated as a local-
ized object prediction task to enable unified learning across
views. On the held-out test set (48 frontal / 58 lateral
/ 15 basal), we report view-wise performance using per-
image mean NME and PCK with view-adaptive normaliza-
tion (alL–alR for frontal/basal; tr–gn with n–gn fallback for
lateral). The model achieves NME median (IQR) of 0.0637
(0.0479) for frontal, 0.0148 (0.0275) for lateral, and 0.0302
(0.0246) for basal, with PCK@0.10 of 85.76%, 91.74%, and
94.67%, respectively. These results support practical multi-
view facial–nasal landmarking for landmark-driven rhino-
plasty analysis and planning.