Breast cancer is one of the deadliest forms of cancer in women but the disease has a good prognosis when diagnosed early. The gold standard for the diagnosis of breast cancer is mammography imaging analysis but the acquisition of mammograms is a painful and embarrassing procedure for women involving breast compression. Alternative methods have been investigated in the last years, including breast thermography, which does not involve ionizing radiation, pain or contact with the breast. However, the accuracy of these modern techniques still needs to be improved to allow the widespread use in practical applications but machine learning techniques have brought in an increased accuracy and reduction in false positives and false negatives to the analysis of breast thermograms. We propose a methodology for detecting and classifying breast lesions using a database of real images of Brazilian patients. We divide our methodology into three steps. In the first step, the shape and texture characteristics of breast thermograms are extracted using Zernike and Haralick moments. The second step is the feature selection process using multi-objective binary optimization algorithms based on swarm intelligence. The third step is to analyze the best vectors’ classification using eleven algorithms such as Convolutional Neural Networks, Extreme Learning Machines, and Support Vector Machines. Finally, we discuss the computational time and performance of various techniques based on swarm intelligence, artificial neural networks, and statistical models to improve the computational time and accuracy of breast cancer diagnoses. Indeed, we observe that the feature selection process has helped us decrease computational time with a high potential to improve diagnostic accuracy. We also demonstrate that the extracted features considering the shape of breast lesions are highly important to a high diagnostic accuracy.

乳腺癌是女性最致命的癌症形式之一，但如果及早诊断，这种疾病的预后良好。诊断乳腺癌的金标准是乳房 X 线摄影成像分析，但对于涉及乳房压迫的女性来说，获取乳房 X 线照片是一个痛苦和尴尬的过程。过去几年已经研究了替代方法，包括乳房热成像，它不涉及电离辐射、疼痛或与乳房的接触。然而，这些现代技术的准确性仍然需要提高，以允许在实际应用中广泛使用，但机器学习技术已经提高了准确性，并减少了乳房温度图分析的误报和漏报。我们提出了一种使用巴西患者真实图像数据库检测和分类乳房病变的方法。我们将我们的方法分为三个步骤。第一步，使用 Zernike 和 Haralick 矩提取乳房热像图的形状和纹理特征。第二步是使用基于群智能的多目标二元优化算法的特征选择过程。第三步是使用卷积神经网络、极限学习机和支持向量机等十一种算法来分析最佳向量的分类。最后，我们讨论了基于群智能、人工神经网络和统计模型的各种技术的计算时间和性能，以提高乳腺癌诊断的计算时间和准确性。确实，我们观察到特征选择过程帮助我们减少了计算时间，并具有提高诊断准确性的巨大潜力。我们还证明，考虑到乳房病变形状的提取特征对于高诊断准确性非常重要。