ABSTRACT

Infrared thermography is a technique that can detect anomalies in temperature patterns which can indicate some breast pathologies including breast cancer. One limitation of the method is the absence of standardised thermography interpretation procedures. Deep learning models have been used for pattern recognition and classification of objects and have been adopted as an adjunct methodology in medical imaging diagnosis. In this paper, the use of a deep convolutional neural network (CNN) with transfer learning is proposed to automatically classify thermograms into two classes (normal and abnormal). A population of 311 female subjects was considered analysing two approaches to test the CNN’s performance: one with a balanced class distribution and the second study in a typical screening cohort, with a low prevalence of abnormal thermograms. Results showed that the transfer-learned ResNet-101 model had a sensitivity of 92.3% and a specificity of 53.8%, while with an unbalanced distribution the values were 84.6% and 65.3%, respectively. These results suggest that the model presented in this work can classify abnormal thermograms with high sensitivity which validates the use of infrared thermography as an adjunct method for breast cancer screening.

摘要

红外热成像是一种可以检测温度模式异常的技术，这些异常可以指示一些乳腺疾病，包括乳腺癌。该方法的一个限制是缺乏标准化的热成像解释程序。深度学习模型已用于对象的模式识别和分类，并已被用作医学影像诊断的辅助方法。在本文中，提出了使用带有迁移学习的深度卷积神经网络 (CNN) 来自动将热谱图分为两类（正常和异常）。311 名女性受试者被考虑分析两种方法来测试 CNN 的性能：一种具有平衡的类别分布，另一种是在典型的筛查队列中进行的研究，异常热谱图的发生率较低。结果表明，迁移学习的 ResNet-101 模型的敏感性为 92.3%，特异性为 53.8%，而在不平衡分布的情况下，该值分别为 84.6% 和 65.3%。这些结果表明，这项工作中提出的模型可以对异常热像图进行高灵敏度分类，这验证了红外热像图作为乳腺癌筛查的辅助方法的使用。