arXiv 2509.128182025-09-16cs.CVcs.AI

放射学基础模型的数据缩放定律

Data Scaling Laws for Radiology Foundation Models

Microsoft Health Futures UK（微软健康未来英国）；Microsoft Health Futures US（微软健康未来美国）；Mayo Clinic M.D（梅奥诊所）；Radiology AI Lab, Mayo Clinic M.D（放射学AI实验室，梅奥诊所）；Microsoft Health & Life Sciences（微软健康与生命科学）；Cambridge University Hospitals M.D（剑桥大学医院）

Maximilian Ilse, Harshita Sharma, Anton Schwaighofer, Sam Bond-Taylor, Fernando Pérez-García, Olesya Melnichenko, Anne-Marie G. Sykes, Kelly K. Horst, Ashish Kh… 展开作者

Maximilian Ilse, Harshita Sharma, Anton Schwaighofer, Sam Bond-Taylor, Fernando Pérez-García, Olesya Melnichenko, Anne-Marie G. Sykes, Kelly K. Horst, Ashish Khandelwal, Maxwell Reynolds, Maria T. Wetscherek, Noel C. F. Codella, Javier Alvarez-Valle, Korfiatis Panagiotis, Valentina Salvatelli

阅读 PDF

arXiv 页面 ↗HTML 原文 ↗

AI总结：

本研究系统评估了MI2和RAD-DINO在350万张胸片上的持续预训练，发现两者在发现和管路任务上各具优势，结构化监督可提升性能，仅需3万域内样本即可超越开放权重模型，证实了中心特定预训练的价值。

摘要：

诸如 CLIP 和 DINOv2 等基于网络规模数据训练的基础视觉编码器，在跨任务和数据集上展现出强大的迁移性能。然而，医学影像基础模型仍受限于较小的数据集，这限制了我们对数据规模和预训练范式如何影响该场景下性能的理解。在本研究中，我们系统地研究了两种视觉编码器 MedImageInsight (MI2) 和 RAD-DINO（分别代表 CLIP 和 DINOv2 两种主要编码器范式）在来自单一机构的最多 3.5M 张胸部 X 光片上的持续预训练，并保持计算资源和评估协议不变。我们在分类（放射学发现、管路）、分割（管路）和放射学报告生成任务上进行评估。尽管先前的工作主要集中在与放射学发现相关的任务上，但我们引入了管路任务以平衡这种偏差，并评估模型提取沿细长结构保持连续性特征的能力。实验表明，MI2 在发现相关任务上扩展更有效，而 RAD-DINO 在管路相关任务上表现更强。令人惊讶的是，使用 UniCL 结合报告和结构化标签持续预训练 MI2 能提升性能，凸显了大规模结构化监督的价值。我们进一步表明，对于某些任务，仅需 30k 域内样本即可超越开放权重基础模型。这些结果突显了特定中心持续预训练的实用性，使医疗机构能利用域内数据获得显著性能提升。

英文摘要：

Foundation vision encoders such as CLIP and DINOv2, trained on web-scale data, exhibit strong transfer performance across tasks and datasets. However, medical imaging foundation models remain constrained by smaller datasets, limiting our understanding of how data scale and pretraining paradigms affect performance in this setting. In this work, we systematically study continual pretraining of two vision encoders, MedImageInsight (MI2) and RAD-DINO representing the two major encoder paradigms CLIP and DINOv2, on up to 3.5M chest x-rays from a single institution, holding compute and evaluation protocols constant. We evaluate on classification (radiology findings, lines and tubes), segmentation (lines and tubes), and radiology report generation. While prior work has primarily focused on tasks related to radiology findings, we include lines and tubes tasks to counterbalance this bias and evaluate a model's ability to extract features that preserve continuity along elongated structures. Our experiments show that MI2 scales more effectively for finding-related tasks, while RAD-DINO is stronger on tube-related tasks. Surprisingly, continually pretraining MI2 with both reports and structured labels using UniCL improves performance, underscoring the value of structured supervision at scale. We further show that for some tasks, as few as 30k in-domain samples are sufficient to surpass open-weights foundation models. These results highlight the utility of center-specific continual pretraining, enabling medical institutions to derive significant performance gains by utilizing in-domain data.

机构、总结、中文摘要均由AI分析而来，可能存在错误，请以论文正文为准。