JMIR Publications

ABSTRACT

Background:

Rhinoplasty is among the most technically demanding procedures in facial plastic and reconstructive surgery, requiring simultaneous management of cartilage conservation, structural integrity, donor site selection, and patient-specific aesthetic goals. Surgical planning relies predominantly on surgeon experience and informal estimation, with no standardized quantitative framework for comparing surgical complexity across candidate plans. Existing AI approaches to surgical planning more broadly lack explicit mathematical representations of anatomic constraints and tissue redistribution logic, limiting their ability to produce structurally feasible, interpretable plans.

Objective:

We propose a proof-of-concept AI planning framework based on unbalanced optimal transport (UOT), introducing transport cost as a novel AI-derived surgical complexity index. We use rhinoplasty cartilage redistribution planning as a motivating domain, with the framework generalizable to other surgical domains involving redistribution of biological material under structural constraints.

Methods:

Pre-surgical and target cartilage anatomy are represented as probability distributions across five anatomic regions. A cost matrix encodes surgical difficulty of redistribution, solved via majorization-minimization with Kullback-Leibler divergence. A two-layer cost matrix incorporates an indication layer reducing donor site costs when septal supply falls below a threshold encoding the L-strut constraint. Two synthetic scenarios are evaluated: primary and revision rhinoplasty with varying septal reserve. Face validity was assessed by comparing framework-derived transport costs against complexity ratings from six fellowship-trained facial plastic and reconstructive surgeons, using Spearman correlation and the Jonckheere-Terpstra test for ordered trend.

Results:

In the primary scenario, UOT costs increased monotonically across conservative, moderate, and aggressive goals (0.000, 0.038, 0.127); ear recruitment occurred exclusively in the aggressive plan. In the revision scenario, rib harvest occurred across all plans (0.088 to 0.130). Marginal cost analysis identified the conservative-to-moderate transition as the principal complexity driver. Transport costs showed strong concordance with expert ratings (Spearman rho = 0.943, P = .005), with a statistically significant monotonic trend confirmed in both scenarios (P < .01).

Conclusions:

UOT provides a generalizable mathematical foundation for constraint-aware AI surgical planning, applicable to any surgical problem where anatomic states, redistribution costs, and indication rules can be specified. Transport cost serves as a domain-general, AI-derived complexity index, supported here by preliminary face validity against expert ratings, pending calibration against clinical outcomes data.