Lago Miguel A, Badano Aldo
U.S. Food and Drug Administration, Silver Spring, Maryland, USA.
Med Phys. 2025 Mar;52(3):1960-1968. doi: 10.1002/mp.17571. Epub 2024 Dec 17.
In silico clinical trials are becoming more sophisticated and allow for realistic assessment and comparisons of medical image system models. These fully computational models enable fast and affordable trial designs that can closely capture trends seen on real clinical trials.
To evaluate three breast imaging system models for digital mammography (DM) and digital breast tomosynthesis (DBT) in a fully-in-silico longitudinal study.
We developed in silico models for three different breast imaging systems by modeling relevant characteristics such as detector technology, pixel size, number of projections, and angular span. We use a computational image reader to detect masses at different growing stages to compute the relative system performance. Similarly, we compare calcification cluster detectability across systems. The Detectability area under the ROC curve (AUC) was calculated for each combination of breast density, device model, lesion size and type, and search area. We compared the absolute and relative AUC values for DM and DBT. The trial consisted of 45 000 simulated images corresponding to 750 virtual digital patient models.
We observed proportional AUC values with increasing mass size. On the other hand, higher breast densities showed lower AUC values. For masses, we found significant performance differences between device models. The highest average AUC difference between DBT and DM was 0.109, benefiting DBT. For calcifications, DM showed higher performance than DBT, especially in highly dense breasts. The highest AUC difference on a model was -0.055, benefiting DM.
In this fully-in-silico imaging trial, we compared three imaging systems with different detector technologies on the same cohort of virtual digital patient models. We found that breast device systems can lead to visibility differences in masses and calcifications. Our longitudinal, multi-device in silico study was possible because of the versatility and flexibility of in silico methods. This study shows the advantages of this in silico methodology in lowering the resources needed for device development, optimization, and regulatory evaluation.
计算机模拟临床试验正变得越来越复杂，能够对医学影像系统模型进行实际评估和比较。这些完全基于计算的模型能够实现快速且经济的试验设计，从而可以紧密捕捉真实临床试验中出现的趋势。
在一项完全计算机模拟的纵向研究中评估三种用于数字乳腺摄影（DM）和数字乳腺断层合成（DBT）的乳腺成像系统模型。
我们通过对探测器技术、像素大小、投影数量和角度跨度等相关特征进行建模，开发了三种不同乳腺成像系统的计算机模拟模型。我们使用一个计算图像读取器来检测处于不同生长阶段的肿块，以计算相对系统性能。同样，我们比较了不同系统对钙化簇的可检测性。针对乳腺密度、设备模型、病变大小和类型以及搜索区域的每种组合，计算ROC曲线下的可检测性面积（AUC）。我们比较了DM和DBT的绝对和相对AUC值。该试验由对应于750个虚拟数字患者模型的45000张模拟图像组成。
我们观察到随着肿块大小增加，AUC值成比例变化。另一方面，较高的乳腺密度显示出较低的AUC值。对于肿块，我们发现不同设备模型之间存在显著的性能差异。DBT和DM之间的最高平均AUC差异为0.109，DBT更具优势。对于钙化，DM表现出比DBT更高的性能，尤其是在高密度乳腺中。一个模型上的最高AUC差异为 -0.055，DM更具优势。
在这项完全计算机模拟的成像试验中，我们在同一组虚拟数字患者模型上比较了三种具有不同探测器技术的成像系统。我们发现乳腺设备系统会导致在肿块和钙化的可见性上存在差异。由于计算机模拟方法的通用性和灵活性，我们的纵向、多设备计算机模拟研究得以实现。这项研究展示了这种计算机模拟方法在降低设备开发、优化和监管评估所需资源方面的优势。
