PelletRayTion: unsupervised contaminants detection in pelleted seeds using high-throughput tomography and 3D β-VAE

Seed coating improves agricultural handling and germination, but obscures visual identification and complicates quality control and regulatory compliance. Conventional de-coating is destructive, time-intensive and raises concerns about reliability and analyst safety. To compensate for these limitations, we introduce PelletRayTion , a non-destructive, automated alternative for contaminants detection in pelleted seeds using high-throughput tomography and unsupervised machine learning. Our fully unsupervised pipeline for contamination detection using β-VAE was trained exclusively on pure tomography data. This approach was evaluated on Beta vulgaris and Cichorium endivia , artificially contaminated with three other species, and reciprocally Cichorium endivia or Beta vulgaris . Our models reached sufficient generalisation with a data size of 800 volumes (64×64×64) for Beta vulgaris and 700 (64×64×64) for Cichorium endivia, using latent space dimensionality of 512 and 256, respectively. Across 1–50% contamination, they achieved balanced accuracy of 98.45 ± 0.95% ( Beta vulgaris ) and 97.47 ± 0.95% ( Cichorium endivia ), with 0% false negatives and low false alerts (3.10% − 5.06%). In conclusion, these results demonstrate that PelletRayTion provides a rapid, safe, and highly accurate alternative to conventional pelleted seed testing. This approach eliminated the need for massive annotated datasets while maintaining strong performance across different species and contaminants. It holds a significant potential for routine testing, biosecurity and regulatory workflows.