[Paper Review] AutoFFS: Adversarial Deformations for Facial Feminization Surgery Planning

Paul Friedrich, Florentin Bieder+2 more|arXiv (Cornell University)|Mar 2, 2026

Facial Rejuvenation and Surgery Techniques0 citations

TL;DR

AutoFFS generates counterfactual skull morphologies by applying deformation-based targeted adversarial attacks on an ensemble of binary sex classifiers, to aid quantitative planning for facial feminization surgery.

ABSTRACT

Facial feminization surgery (FFS) is a key component of gender affirmation for transgender and gender diverse patients, aiming to reshape craniofacial structures toward a female morphology. Current surgical planning procedures largely rely on subjective clinical assessment, lacking quantitative and reproducible anatomical guidance. We therefore propose AutoFFS, a novel data-driven framework that generates counterfactual skull morphologies through adversarial free-form deformations. Our method performs a deformation-based targeted adversarial attack on an ensemble of pre-trained binary sex classifiers that learned sexual dimorphism, effectively transforming individual skull shapes toward the target sex. The generated counterfactual skull morphologies provide a quantitative foundation for preoperative planning in FFS, driving advances in this largely overlooked patient group. We validate our approach through classifier-based evaluation and a human perceptual study, confirming that the generated morphologies exhibit target sex characteristics.

Motivation & Objective

• Motivate quantitative, reproducible guidance for facial feminization surgery (FFS) planning beyond subjective assessments.
• Develop a data-driven framework to create counterfactual skull morphologies toward the opposite sex.
• Leverage deformation-based adversarial attacks on sex classifiers to produce target-sex skull shapes.
• Provide a quantitative foundation for personalized surgical planning grounded in anatomical data.

Proposed method

• Train an ensemble of binary sex classifiers to capture sexual dimorphism in skull morphology.
• Parameterize deformations using a 3D cubic B-spline free-form deformation (FFD) with a controllable lattice.
• Perform a deformation-based targeted adversarial attack at test time by optimizing control point offsets to maximize target-sex probability.
• Regularize deformations with smoothness and bending energy terms to ensure realistic, plausible transformations.
• Use a smooth worst-case margin loss over the ensemble (and mirrored flips) to drive all classifiers past a margin γ.
• Backpropagate through the deformation to update control points via Adam optimization.

Experimental results

Research questions

• RQ1 Can counterfactual skull morphologies toward a target sex be generated in a data-driven, anatomically plausible manner?
• RQ2 Do ensemble-based adversarial deformations produce morphologies that align with target-sex characteristics across diverse skull shapes?
• RQ3 Are the generated morphologies perceptually aligned with the intended sex, and can they offer objective guidance for surgical planning?

Key findings

• classifiers trained on skull MR-derived bone morphology achieve high hold-out performance (Acc above 0.85, AUROC > 0.93 across architectures).
• Deformations guided by the ensemble reliably push skull morphology toward the target sex in evaluation classifiers and mirrored flips.
• Visual analysis shows largest deformations in chin, brow ridges, forehead, and cheekbones, consistent with sex-dimorphic features.
• A perceptual study shows real skulls are correctly classified ~81% of the time, while transformed skulls are perceived as the target sex in ~63% of cases.
• The ensembling strategy improves consistency and robustness of the deformation outcomes compared to using a single model.