Thermal Modeling of Patient-Specific Breast Cancer With Physics-Based Artificial Intelligence
Thermal Modeling of Patient-Specific Breast Cancer With Physics-Based Artificial Intelligence
复制标题
利用基于物理的人工智能对患者特异性乳腺癌进行热建模
DOI：
10.1115/1.4055347
复制
发表时间：
2023
期刊：
影响因子：
--
通讯作者：
Kandlikar, S. G.
中科院分区：
文献类型：
--
作者：
Perez-Raya, I.;Kandlikar, S. G.
乳腺癌是女性中常见的癌症。它与由于肿瘤中更高的代谢和血管生成导致的血管增加而产生的热量增加有关。热变化导致乳房表面温度曲线的变化。红外成像是FDA批准的乳房X光检查的替代品，后者利用表面温度的变化来检测癌症。为了将红外成像应用于临床，有必要开发有效的技术来模拟肿瘤特征与乳房表面温度之间的关系。目前的工作描述了乳腺癌的热建模与物理信息的神经网络。基于边界条件和应满足的控制方程，将损耗分配给域中的随机点。TensorFlow中的Adam优化器将损失最小化，以找到满足边界条件和生物热方程的温度场或热导率。反向传播计算生物热方程中的导数。对三个患者特定病例的分析表明，机器学习模型准确地再现了ansys-fluent模拟给出的热行为。此外，良好的协议之间的模型预测和红外图像观察。此外，神经网络准确地恢复导热系数在6.5%的相对误差。
Breast cancer is a prevalent form of cancer among women. It is associated with increased heat generation due to higher metabolism in the tumor and increased blood vessels resulting from angiogenesis. The thermal alterations result in a change in the breast surface temperature profile. Infrared imaging is an FDA-approved adjunctive to mammography, which employs the surface temperature alterations in detecting cancer. To apply infrared imaging in clinical settings, it is necessary to develop effective techniques to model the relation between the tumor characteristics and the breast surface temperatures. The present work describes the thermal modeling of breast cancer with physics-informed neural networks. Losses are assigned to random points in the domain based on the boundary conditions and governing equations that should be satisfied. The Adam optimizer in TensorFlow minimizes the losses to find the temperature field or thermal conductivity that satisfies the boundary conditions and the bioheat equation. Backpropagation computes the derivatives in the bioheat equation. Analyses of the three patient-specific cases show that the machine-learning model accurately reproduces the thermal behavior given byansys-fluentsimulation. Also, good agreement between the model prediction and the infrared images is observed. Moreover, the neural network accurately recovers the thermal conductivity within 6.5% relative error.