Supervised Bayesian learning for breast cancer detection in terahertz imaging
Supervised Bayesian learning for breast cancer detection in terahertz imaging
复制标题
DOI：
10.1016/j.bspc.2021.102949
复制
发表时间：
2021-07-22
影响因子：
5.1
通讯作者：
Wu, Jingxian
中科院分区：
文献类型：
--
作者：
Chavez, Tanny;Vohra, Nagma;Wu, Jingxian
本文提出了一种监督多项式贝叶斯学习算法乳腺癌检测使用太赫兹（THz）成像的新鲜切除的小鼠肿瘤。该算法利用多项贝叶斯概率回归方法，通过使用两种不同的模型，多项式回归模型和核回归模型，建立太赫兹数据和分类结果之间的联系。这种基于模型的学习方法只使用少量的模型参数，因此与其他深度学习方法相比，它需要更少的训练数据。通过使用福尔马林固定的石蜡包埋（FFPE）样本的组织病理学结果作为基础事实来执行算法的训练阶段。由于样品脱水，新鲜切除的样品与其FFPE对应物之间通常存在相当大的形状失配，并且这种失配对训练数据的质量产生负面影响。我们建议通过使用一种创新的基于可靠性的训练数据选择方法来解决这一挑战，其中训练数据的可靠性通过使用具有软概率输出的无监督期望最大化（EM）分类算法来量化和估计。实验结果表明，提出的多项贝叶斯概率回归模型与可靠性为基础的训练数据选择取得了更好的性能比现有的方法。总的来说，这些结果表明，建议的监督分割模型代表了一种很有前途的技术，该技术的区域检测与太赫兹成像的新鲜切除的乳腺癌样本。
This paper proposes a supervised multinomial Bayesian learning algorithm for breast cancer detection using terahertz (THz) imaging of freshly excised murine tumors. The proposed algorithm utilizes a multinomial Bayesian probit regression approach, which establishes the link between THz data and classification results by using two different models, a polynomial regression model and a kernel regression model. Such a model-based learning approach employs only a small number of model parameters, thus it requires much less training data when compared with alternative deep learning methods. The training phase of the algorithm is performed by using the histopathology results of formalin-fixed, paraffin embedded (FFPE) samples as ground truth. There is usually a considerable shape mismatch between the freshly excised sample and its FFPE counterpart due to sample dehydration, and such mismatch negatively impacts the quality of the training data. We propose to address this challenge by using an innovative reliability-based training data selection method, where the reliability of the training data is quantified and estimated by using an unsupervised expectation maximization (EM) classification algorithm with soft probabilistic output. Experiment results demonstrate that the proposed multinomial Bayesian probit regression models with reliability-based training data selection achieve better performance than existing methods. Overall, these results demonstrate that the proposed supervised segmentation models represent a promising technique for the region detection with THz imaging of freshly excised breast cancer samples.