Researchers have developed BrainAnytime, a novel pretraining framework designed for brain image analysis that can handle incomplete or varied imaging data. This unified model accepts any available imaging sequences, from single MRI scans to multimodal MRI and PET data. BrainAnytime leverages cross-modal distillation and atlas-guided masking within a shared 3D masked autoencoder to learn structural-molecular correspondences. In evaluations across four downstream tasks and five clinical modality settings, BrainAnytime demonstrated superior performance compared to modality-specific models and existing baselines, notably achieving significant accuracy improvements in classifying cognitive states. AI
IMPACT Enables more robust AI diagnostics for brain conditions by accommodating real-world, incomplete medical imaging data.
RANK_REASON Publication of a new AI framework and associated paper on arXiv. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
