Large-scale chest x-ray datasets have been curated for the detection of abnormalities using deep learning, with the potential to provide substantial benefits across many clinical applications. However, each dataset focuses only on detecting a subset of findings that can be simultaneously present in a patient, thereby limiting its clinical utility. Therefore, data harmonization is crucial to leverage these datasets in aggregate to train clinically-useful, robust models with a complete representation of all abnormalities that may occur within the thorax. To that end, we propose surgical aggregation, a collaborative learning framework for harmonizing and aggregating knowledge from distributed heterogeneous datasets with partial disease annotations. We evaluate surgical aggregation across synthetic iid datasets and real-world large-scale non-iid datasets with partial annotations. Our results indicate that surgical aggregation significantly outperforms current strategies, has better generalizability, and has the potential to revolutionize the development clinically-useful models as AI-assisted disease characterization becomes a mainstay in radiology.

翻译：大规模胸部X光数据集已被整理用于基于深度学习的异常检测，有望在众多临床应用中带来显著收益。然而，每个数据集仅专注于检测患者体内可能同时存在的部分病变，从而限制了其临床实用性。因此，数据协调对于聚合利用这些数据集至关重要，旨在训练具有完整胸腔异常表征的临床实用且鲁棒的模型。为此，我们提出手术聚合这一协同学习框架，用于协调和整合具有部分疾病标注的分布式异构数据集中的知识。我们在合成的独立同分布数据集和真实世界的大规模非独立同分布部分标注数据集上对手术聚合进行了评估。结果表明，手术聚合在性能上显著优于现有策略，具有更强的泛化能力，并有望随着人工智能辅助疾病诊断成为放射学主流，彻底改变临床实用模型的开发范式。