Accurate identification of extramural vascular invasion (EVI) and mesorectal fascia invasion (MFI) on MRI is critical for rectal cancer risk stratification and treatment planning. However, visual assessment remains subjective and can vary across radiologists, scanners and institutions.

This multicentre study developed an AI framework combining pre-trained foundation models, MRI harmonization and multiplanar imaging to automatically classify EVI and MFI from pre-treatment T2-weighted MRI in 331 patients across three European hospitals.

For EVI, combining axial and sagittal MRI information achieved the best performance, with an AUC of 0.82, compared with 0.74 for the CHAIMELEON benchmark. For MFI, the highest performance was obtained from harmonized axial images (AUC 0.77 vs 0.75 benchmark). Foundation model-derived features also outperformed conventional CNN approaches.

Importantly, the study shows that harmonization and multiplanar fusion should be tailored to the specific imaging biomarker: harmonization improved MFI classification but reduced EVI performance, while multiplanar information particularly benefited EVI detection.

As one of the first studies to automate both EVI and MFI assessment across axial and sagittal MRI, this work highlights the potential of foundation models to deliver more consistent, generalizable and scalable AI-assisted rectal cancer staging across centres.