PURPOSE: Quantitative analysis of craniofacial morphology in animal models is critical for studying normal development and the pathogenesis of congenital anomalies. Conventional morphometric approaches rely on manual landmarking, which is labor-intensive, biased toward predefined features, and difficult to reproduce across laboratories. The objective of this work was to evaluate the feasibility of unsupervised machine learning methods for automated, reproducible cranial phenotyping using high-resolution imaging data. METHODS: Micro-CT scans of mouse skulls were obtained from an open source data repository, and divided based on strain, sex, and developmental age. A preprocessing pipeline was implemented that included segmentation of the facial skeleton and cranial vault, alignment of scans to a standardized orientation, and cropping to isolate craniofacial structures. A SwinUNETR deep learning architecture was trained on a reduced dataset of 500 skulls, and downsampled to 0.1 mm spacing, to perform automated skeletal segmentation. Following segmentation, automated landmarking and principal component analysis were used to generate low-dimensional shape descriptors capturing morphological variation across specimens without reliance on predefined phenotypic categories. RESULTS: The segmentation model achieved a Dice similarity coefficient of 77%, indicating reliable separation of craniofacial structures despite the reduced dataset size and resolution. Integration of segmentation with PCA produced quantitative shape descriptors capable of capturing both gross morphological differences and subtle structural variations across developmental stages and genetic backgrounds. Using overlaid vectors approximating the changes observed in orientation and size allowed for a user-friendly engagement with the computational output of the model. CONCLUSION: Unsupervised machine learning provides a scalable and unbiased framework for automated cranial phenotyping. By eliminating dependence on manual landmarking and predefined categories, this approach enables reproducible, high-throughput morphometric analysis suitable for large datasets and multi-laboratory studies. Although developed and tested in mice, the methodology is adaptable to other model organisms, and is likely to lower the threshold for non-craniofacial laboratories to assess the craniofacial phenotypes of the genetic variants they breed over time. *Source: https://ps-rc.org/meeting/Program/2026/CS51.cgi*

Abstract ID: CS51