Detection and classification of breast lesions in thermographic images by deep transfer learning hybrid architectures

This paper aims to investigate the potential of thermography as a complementary technique for breast cancer detection. To reach this objective, we propose a hybrid architecture founded on deep transfer learning for feature extraction, coupled with shallow algorithms for decision-making, targeting the detection and classification of breast lesions in thermographic images.

The study employs convolutional neural networks (CNNs) to classify breast cancer using thermographic images. Two approaches are compared: one utilizing original feature vectors extracted by CNNs, and the other using vectors reduced by Particle Swarm Optimization (PSO) for feature selection.

The results highlight the effectiveness of both approaches. For complete feature vectors, the highest accuracy achieved was 79.92% with the Inception V3 CNN combined with SVM using a 3rd-degree polynomial kernel. The best sensitivity and specificity values were achieved by the Inception V3 CNN combined with SVM using an RBF kernel with γ = 0.25, yielding sensitivity of 100% and specificity of 99.49%. The highest AUC was obtained by the same combination with an AUC of 0.83. With PSO-selected features, the highest accuracy was 78.55% using the Inception V3 CNN combined with SVM using a 4th-degree polynomial kernel. The PSO-based approach reduced feature dimensions and training time without significant loss of accuracy.

In conclusion, the study demonstrates the potential of thermography combined with deep learning techniques for accurate breast cancer detection. The PSO-based feature selection approach provides insights into reducing feature dimensions and training time. These findings contribute to the integration of thermography and advanced machine learning for enhanced breast cancer diagnosis.