Deep learning-based 3D forensic facial reconstruction: a volumetric morphological inference approach using a skull-conditioned U-Net

Forensic facial reconstruction (FFR) is an investigative aid used in cases where primary identifiers such as DNA, dental records, or fingerprints are unavailable or inconclusive. Rather than establishing identity, FFR aims to generate an anatomically plausible facial approximation that may facilitate recognition and help narrow the pool of potential candidates. Reconstructing facial soft tissue from cranial geometry remains an inherently ill-posed inverse problem, as the skull does not uniquely determine the overlying facial morphology. In this study, skull-conditioned facial reconstruction is formulated as a volumetric morphological inference task using a fully automated deep learning framework. Given the limited dataset size, this work is explicitly presented as a feasibility-focused concept study designed to evaluate the methodological viability of skull-conditioned volumetric reconstruction and to characterize anatomically structured reconstruction errors, rather than to establish population-level generalizability. Based on craniofacial CT data, a 3D U-Net learns a voxel-to-voxel mapping from binary skull volumes to facial soft-tissue envelopes. The model relies exclusively on cranial geometry without incorporating demographic metadata, thereby isolating anatomical information encoded in the skull and avoiding assumptions often unavailable in forensic contexts. A composite loss function is employed to address volumetric class imbalance and promote stable reconstruction of continuous soft-tissue structures. Reconstruction accuracy is evaluated using volumetric overlap and surface-based distance metrics. On an independent test set, the model achieves a mean Dice coefficient of 0.93 ± 0.01 and a mean 95th -percentile Hausdorff distance (HD95) of 2.8 ± 0.8 mm. Spatial error analysis indicates higher accuracy in skull-anchored regions and larger deviations in anatomically under-constrained areas, consistent with known principles of anatomical constraint. Within this feasibility-focused scope, these findings provide preliminary evidence that skull-conditioned volumetric deep learning can generate anatomically coherent facial approximations while also revealing the intrinsic limitations of reconstruction in anatomically under-constrained regions.