Background/Objectives: We provide a framework for determining how far into the future the spatiotemporal dynamics of tumor growth can be accurately predicted using routinely available magnetic resonance imaging (MRI) data. Our analysis is applied to a coupled set of reaction-diffusion equations describing the spatiotemporal development of tumor cellularity and vascularity, initialized and constrained with diffusion-weighted (DW) and dynamic contrast-enhanced (DCE) MRI data, respectively. Methods: Motivated by experimentally acquired murine glioma data, the rat brain serves as the computational domain within which we seed an in silico tumor. We generate a set of 13 virtual tumors defined by different combinations of model parameters. The first parameter combination was selected as it generated a tumor with a necrotic core during our simulated ten-day experiment. We then tested 12 additional parameter combinations to study a range of high and low tumor cell proliferation and diffusion values. Each tumor is grown for ten days via our model system to establish “ground truth” spatiotemporal tumor dynamics with an infinite signal-to-noise ratio (SNR). We then systematically reduce the quality of the imaging data by decreasing the SNR, downsampling the spatial resolution (SR), and decreasing the sampling frequency, our proxy for reduced temporal resolution (TR). With each decrement in image quality, we assess the accuracy of the calibration and subsequent prediction by comparing it to the corresponding ground truth data using the concordance correlation coefficient (CCC) for both tumor and vasculature volume fractions, as well as the Dice similarity coefficient for tumor volume fraction. Results: All tumor CCC and Dice scores for each of the 13 virtual tumors are >0.9 regardless of the SNR/SR/TR combination. Vasculature CCC scores with any SR/TR combination are >0.9 provided the SNR ≥ 80 for all virtual tumors; for the special case of high-proliferating tumors (i.e., proliferation > 0.0263 day−1), any SR/TR combination yields CCC and Dice scores > 0.9 provided the SNR ≥ 40. Conclusions: Our systematic evaluation demonstrates that reaction-diffusion models can maintain acceptable longitudinal prediction accuracy—especially for tumor predictions—despite limitations in the quality and quantity of experimental data.
背景/目的：本研究建立了一个评估框架，用于确定利用常规磁共振成像（MRI）数据能够准确预测肿瘤生长时空动态的未来时间范围。我们采用一组描述肿瘤细胞密度与血管密度时空演化的反应-扩散耦合方程进行分析，并分别通过扩散加权（DW）和动态对比增强（DCE）MRI数据进行模型初始化与参数约束。方法：基于实验获取的小鼠胶质瘤数据，我们以大鼠脑组织作为计算域植入虚拟肿瘤。通过13组不同模型参数组合生成虚拟肿瘤集，其中首组参数被选定是因为其在模拟的十天实验中能产生具有坏死核心的肿瘤。随后测试的12组参数组合用于研究高/低肿瘤细胞增殖率与扩散系数的变化范围。通过模型系统使每个肿瘤生长十天，建立具有无限信噪比（SNR）的“真实基准”时空肿瘤动态。通过逐步降低SNR、下采样空间分辨率（SR）以及减少采样频率（作为时间分辨率降低的代理变量），我们系统性地模拟了影像数据质量下降的情况。针对每次影像质量降级，通过计算肿瘤与血管体积分数的和谐相关系数（CCC）以及肿瘤体积分数的戴斯相似系数，将校准与预测结果与对应基准数据进行对比，评估其准确性。结果：在所有13组虚拟肿瘤中，无论SNR/SR/TR如何组合，肿瘤CCC与戴斯系数均>0.9。当SNR≥80时，任何SR/TR组合下的血管CCC分数均>0.9；对于高增殖肿瘤（增殖率>0.0263天−1）的特殊情况，只要SNR≥40，任意SR/TR组合均可获得>0.9的CCC与戴斯分数。结论：我们的系统评估表明，尽管实验数据在质量与数量上存在局限，反应-扩散模型仍能保持可接受的纵向预测准确性，尤其对于肿瘤预测具有稳健性。
肿瘤(癌症)患者之家