Convolutional neural network-based computer-aided diagnosis in Hiesho (cold sensation)
Convolutional neural network-based computer-aided diagnosis in Hiesho (cold sensation)
复制标题
基于卷积神经网络的Hiesho（冷感）计算机辅助诊断
DOI：
10.1016/j.compbiomed.2022.105411
复制
发表时间：
2022-03
影响因子：
7.7
通讯作者：
Tianyi Wang;Masayuki Endo;Y. Ohno;Shima Okada;M. Makikawa
中科院分区：
文献类型：
--
作者：
Tianyi Wang;Masayuki Endo;Y. Ohno;Shima Okada;M. Makikawa
Hiesho（冷感）是一个全球性的健康问题，主要发生在女性身上。患有Hiesho的女性会出现四肢冰冷的感觉，这也与其他慢性疾病有关。然而，Hiesho 的诊断仍然存在争议，因为它依赖于问卷等主观方法。定量和自动的 Hiesho 诊断有望提高诊断准确性并减轻患者和医生的负担。我们之前的研究发现，Hiesho 患者的女性前额和足底之间的温度差异显着，因此认为使用热成像图像训练卷积神经网络 (CNN) 有助于 Hiesho 的计算机辅助诊断 (CAD)。因此，本研究提出了一种基于 CNN 的 Hiesho CAD 系统。来自 46 名受试者（23 名 Hiesho 患者和 23 名健康受试者）的总共 5612 张热成像图像被用来训练 AlexNet，并使用准确度、精确度、灵敏度、特异性和 F1 分数评估所提出的 CNN 模型的性能并与其他基于机器学习的模型进行比较。实验结果表明，所提出的基于 CNN 的 Hiesho CAD 模型在所有评估项目中具有最高的性能（100%）。此外，得出的结论是，热成像图像在区分 Hiesho 方面表现出很高的可行性，并且基于 CNN 的 CAD 在自动 Hiesho 诊断方面表现出较高的准确性和可靠性。
Hiesho (cold sensation) is a worldwide health problem primarily occurring in women. Females who suffered from Hiesho reported cold feeling at the extremities, which was also related to other chronic diseases. However, the diagnosis of Hiesho is still controversial because it depends on subjective approaches such as questionnaires. Quantitative and automatic Hiesho diagnosis is expected to increase diagnostic accuracy and lower the burden on patients and doctors. Following our previous study, which found that the temperature difference between females’ foreheads and plantar soles was significant in Hiesho patients, it was considered that training a convolutional neural network (CNN) with thermographic images can contribute to a computer-aided diagnosis (CAD) for Hiesho. Thus, this study proposes a CNN-based Hiesho CAD system. A total of 5612 thermographic images from 46 subjects (23 Hiesho patients and 23 healthy subjects) were used to train AlexNet, and the performance of the proposed CNN model was evaluated and compared with other machine learning-based models using accuracy, precision, sensitivity, specificity, and F1 score. The experimental results showed that the proposed CNN-based Hiesho CAD model had the highest performance (100%) for all evaluated items. In addition, it was concluded that thermographic images showed high feasibility for discriminating Hiesho, and CNN-based CAD showed high accuracy and reliability for automatic Hiesho diagnosis.