中文迈向临床的脑MRI基础模型：FOMO25挑战赛的发现

脑部MRI临床部署面临数据异质与标注昂贵难题。自监督学习（SSL）利用临床工作流中大量未标记数据预训练基础模型，使其能仅需少量监督适应域外数据。然而，现有基础模型受限于预训练规模小及域内基准测试，限制了临床应用潜力。

arXiv:2604.11679v2 Announce Type: replace Abstract: Clinical deployment of automated brain MRI analysis faces a fundamental challenge: clinical data is heterogeneous and noisy, and high-quality labels are prohibitively costly to obtain. Self-supervised learning (SSL) can address this by leveraging the vast amounts of unlabeled data produced in clinical workflows to train robust \textit{foundation models} that adapt out-of-domain with minimal supervision. However, the development of foundation models for brain MRI has been limited by small pretraining datasets and in-domain benchmarking focused