Forensic anthropology, a vital discipline for identifying human remains in legal and humanitarian contexts, is currently facing a systemic "Paradox of Reference." While Machine Learning (ML) offers unparalleled accuracy for estimating biological profiles (sex, age, stature), its application is hindered by a significant "Data Scarcity Wall." Historical skeletal collections are biologically obsolete due to secular trends, while modern identified collections are finite, geographically fragmented, and strictly constrained by ethical and data protection regulations (GDPR).
The SYNOS project (SYNthetic Osteometric Systems) proposes a radical paradigm shift to resolve this bottleneck: shifting from the exclusive reliance on scarce physical remains to the generation of infinite, synthetic, biologically valid data. The project's core objective is to develop SCAFFOLD (Synthetic Creation and Augmentation Framework For Osteometric Learning Data), a bio-informed generative AI architecture.
Adopting a risk-mitigating "Curriculum Learning" strategy, the framework evolves through three levels of complexity: The first method is Tabular Synthesis, employing Conditional Tabular GANs (ctdGAN) with allometric loss functions to generate valid metric profiles. The second is Geometric Synthesis, utilising Graph Neural Networks (GNN) to predict 3D landmarks. The third is Volumetric Synthesis, adapting 3D-StyleGANs to create high-fidelity pseudo-CT scans.
SYNOS is unique in the way it enforces biological constraints within the AI's learning process. This is done in order to prevent anatomical "hallucinations". The validation process will be rigorous, combining quantitative metrics with a qualitative "Visual Turing Test" performed by forensic experts from partner institutions in Milan and Coimbra, and cross-validation against real clinical data from the AP-HM Health Data Warehouse. Furthermore, to ensure the tools are admissible in court, the project integrates Explainable AI (XAI) using SHAP values, providing transparent, case-specific justifications for every identification decision.
SYNOS aims to democratise forensic AI by establishing the first universal, privacy-preserving synthetic reference database. This will enhance the resolution of complex cases ranging from criminal investigations to the identification of migrants and mass disaster victims.
The SYNOS project (SYNthetic Osteometric Systems) proposes a radical paradigm shift to resolve this bottleneck: shifting from the exclusive reliance on scarce physical remains to the generation of infinite, synthetic, biologically valid data. The project's core objective is to develop SCAFFOLD (Synthetic Creation and Augmentation Framework For Osteometric Learning Data), a bio-informed generative AI architecture.
Adopting a risk-mitigating "Curriculum Learning" strategy, the framework evolves through three levels of complexity: The first method is Tabular Synthesis, employing Conditional Tabular GANs (ctdGAN) with allometric loss functions to generate valid metric profiles. The second is Geometric Synthesis, utilising Graph Neural Networks (GNN) to predict 3D landmarks. The third is Volumetric Synthesis, adapting 3D-StyleGANs to create high-fidelity pseudo-CT scans.
SYNOS is unique in the way it enforces biological constraints within the AI's learning process. This is done in order to prevent anatomical "hallucinations". The validation process will be rigorous, combining quantitative metrics with a qualitative "Visual Turing Test" performed by forensic experts from partner institutions in Milan and Coimbra, and cross-validation against real clinical data from the AP-HM Health Data Warehouse. Furthermore, to ensure the tools are admissible in court, the project integrates Explainable AI (XAI) using SHAP values, providing transparent, case-specific justifications for every identification decision.
SYNOS aims to democratise forensic AI by establishing the first universal, privacy-preserving synthetic reference database. This will enhance the resolution of complex cases ranging from criminal investigations to the identification of migrants and mass disaster victims.
Supervisor
Dr. Pascal Adalian, Laboratoire Anthropologie bio culturelle, Droit, Éthique et Santé (ADES), Aix-Marseille Université
Co-Supervisor
Dr. Christophe Roman, Laboratoire d'Informatique et Systèmes, Aix-Marseille Université
Intersectoral partner
AP-HM - Service de Médecine Légale (Forensic Medicine Department), Marseille, France
International partner
LABANOF - University of Milan, Italy & Biological and forensic anthropology - University of Coimbra, Portugal