Gaussian product-of-experts model robustly segments MS lesions when MRI sequences are missing

A multi-scale Gaussian product-of-experts model integrated into nnU-Net achieved a mean Dice of 0.7029 across 26 missing-sequence configurations for MS lesion segmentation on MSSEG2016, and showed generalizability in zero-shot evaluation on MSLesSeg24 (Dice 0.5566 full sequence).

• The MG-PoE fusion rule combines sequence-specific Gaussian experts via precision weighting, enabling the model to handle any combination of available sequences without subset-specific architectures.
• On MS3SEG, MG-PoE significantly outperformed competing methods in mean Dice and composite score, with comparable 95th percentile Hausdorff distance.
• Zero-shot evaluation on MSLesSeg24 demonstrated best overall and missing-sequence Dice and HD95, suggesting cross-dataset robustness.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.