Classifying malignant pulmonary nodules using biophysics-enhanced artificial intelligence

摘要： 肺癌是美国最常见的癌症死亡原因，估计有14万人死亡 在2020年。虽然已经证明肺部筛查可以降低20%的死亡率，但准确的分类 由于假阳性病例很高，恶性肿瘤的诊断仍然是一个未得到满足的需求。现行分类 通过计算机断层扫描(CT)筛查的方法基于统计预测模型，并且 最近的人工智能。提高恶性肿瘤分类准确率将降低成本 和死亡风险，指导临床决策。在这里，我们提出了一个新的框架来进一步 以生物物理学为基础丰富输入信息，提高现有模型的预测能力 计算模型。所提出的计算模型生成基于以下公式的正交信息 物理定律，因此本质上是人工智能无法学习的。我们建议加强生物物理 信息，因为许多研究表明，肿瘤的进展受到 它们生长的物理微环境。 我们在这里的总体目标是提出肺中的生理机械力作为一种信息性和 在现有输入变量的正交生物标记物，以提高预测水平 肺结节的恶性风险。我们的中心假设是，将基于生物物理学的计算 模型与现有的统计模型和人工智能方法相结合，将提高其预测能力 在对恶性肺结节进行分类时。我们的假设是基于已发表的著作和我们初步的 数据表明，肺内的机械应力与肿瘤的发病率和生长密切相关。通过 结合基于生物物理学的计算模型，我们计划评估改进的分类性能在 Logistic回归模型(AIM 1)和深度卷积神经网络(AIM 2)，作为一个高度预测性的模型。基于生物物理学的计算模型与人工智能的耦合 预测工具将在不增加患者经济和健康负担的情况下提高功率预测。这 这里提出的肺癌分类的新方法，对肺癌的诊断和预后有潜在的好处。 其他肺部病理情况，如慢性阻塞性肺疾病(COPD)、纤维化、机械性 通风，以及肺部的细菌和病毒感染。

SUMMARY: Lung cancer is the most common cause of cancer death in the United States with an estimated 140,000 deaths in 2020. While it has been demonstrated that lung screening reduces the mortality by 20%, accurate classification of malignant tumors remains an unmet need due to high rate of false positive cases. Current classification approaches by means of computed tomography (CT) screenings are based on statistical predictive models, and more recently artificial intelligence. Improving the classification accuracy of malignant tumors will reduce costs and the risk of mortality and guide clinical decision making. Here, we propose a novel framework to further improve the predictive power of current models by enriching the input information with biophysics-based computational models. The proposed computational model generates orthogonal information, which is based on laws of physics, and hence intrinsically unlearnable by artificial intelligence. We propose augmenting biophysical information since numerous studies have demonstrated that the tumor progression is strongly affected by the physical microenvironment in which they grow. Our overall objective here is to propose the physiological mechanical forces in lung as an informative and orthogonal biomarker to the existing input variables in state-of-the-art approaches to improve the prediction of malignancy risk in pulmonary nodules. Our central hypothesis is that coupling a biophysics-based computational model with the existing statistical models and artificial intelligence approaches will improve their prediction power in classifying malignant pulmonary nodules. Our hypothesis is based on published works and our preliminary data that the mechanical stresses in the lung are strongly correlated with both tumor incidence and growth. By coupling biophysics-based computational model, we plan to evaluate the improved classification performance in logistic regression model (Aim 1), as a highly interpretable model, and deep convolutional neural network (Aim 2), as a highly predictive model. Coupling biophysics-based computational model to artificial intelligence predictive tools will improve the prediction of power at no added financial and health burden to the patient. This novel approach, proposed here on lung cancer classification, has potential diagnostic and prognostic benefits in other pathological lung conditions such as chronic obstructive pulmonary disease (COPD), fibrosis, mechanical ventilation, and bacterial and viral infection of the lung.