Mayfield John, Tibbetts Katherine Quesada, Rehman Aziz, Levin Millena, Goltz Dayna, Prakash Neelesh
Department of Radiology, USF Health, Tampa, USA.
USF Health Morsani College of Medicine, Tampa, USA.
J Imaging Inform Med. 2024 Dec 2. doi: 10.1007/s10278-024-01337-x.
Small cohorts of certain disease states are common especially in medical imaging. Despite the growing culture of data sharing, information safety often precludes open sharing of these datasets for creating generalizable machine learning models. To overcome this barrier and maintain proper health information protection, foundational models are rapidly evolving to provide deep learning solutions that have been pretrained on the native feature spaces of the data. Although this has been optimized in Large Language Models (LLMs), there is still a sparsity of foundational models for computer vision tasks. It is in this space that we provide an investigation into pretraining Visual Geometry Group (VGG)-16, Residual Network (ResNet)-50, and Dense Network (DenseNet)-121 on an unrelated dataset of 8500 chest CTs which was subsequently fine-tuned to classify bone mineral density (BMD) in 199 breast cancer patients using the L1 vertebra on CT. These semi-foundational models showed significant improved ternary classification into mild, moderate, and severe demineralization in comparison to ground truth Hounsfield Unit (HU) measurements in trabecular bone with the semi-foundational ResNet50 architecture demonstrating the best relative performance. Specifically, the holdout testing AUC was 0.99 (p-value < 0.05, ANOVA versus no pretraining versus ImageNet transfer learning) and F1-score 0.99 (p-value < 0.05) for the holdout testing set. In this study, the use of a semi-foundational model trained on the native feature space of CT provided improved classification in a completely disparate disease state with different window levels. Future implementation with these models may provide better generalization despite smaller numbers of a disease state to be classified.
某些疾病状态的小队列很常见，尤其是在医学成像中。尽管数据共享文化不断发展，但信息安全往往使这些数据集无法公开共享以创建可推广的机器学习模型。为了克服这一障碍并保持适当的健康信息保护，基础模型正在迅速发展，以提供在数据的原生特征空间上进行预训练的深度学习解决方案。尽管这在大语言模型（LLMs）中已得到优化，但用于计算机视觉任务的基础模型仍然很少。正是在这个领域，我们对预训练视觉几何组（VGG）-16、残差网络（ResNet）-50和密集网络（DenseNet）-121在一个包含8500例胸部CT的不相关数据集上进行了研究，随后对其进行微调，以使用CT上的第一腰椎对199例乳腺癌患者的骨密度（BMD）进行分类。与小梁骨中真实的亨氏单位（HU）测量值相比，这些半基础模型在轻度、中度和重度脱矿质的三元分类中表现出显著改善，其中半基础ResNet50架构表现出最佳的相对性能。具体而言，对于留出测试集，留出测试的AUC为0.99（p值<0.05，方差分析与无预训练与ImageNet迁移学习相比），F1分数为0.99（p值<0.05）。在本研究中，使用在CT的原生特征空间上训练的半基础模型在具有不同窗宽水平的完全不同的疾病状态中提供了更好的分类。尽管要分类的疾病状态数量较少，但这些模型的未来应用可能会提供更好的泛化能力。
