Deep-Wavelets and convolutional neural networks to support breast cancer diagnosis on thermography images

Breast cancer is the most dangerous type of cancer and one of the most lethal for women, both in underdeveloped and central countries. Breast thermography is an emerging imaging technique that can be applied as a complementary procedure for screening breast lesions. However, the low knowledge about the interpretation of these images by mastologists makes it difficult to adopt them in clinical practice. Computer-aided detection (CAD) systems can assist medical professionals in this task. Deep learning techniques have contributed to obtaining good results in the classification of biomedical images in general. In this work, we propose Deep-Wavelet Neural Networks (DWNN), convolutional architectures based on the general theory of Wavelets to extract features from images. For classification of thermographic images, we propose hybrid architectures based on deep network for feature extraction, Random Forests for selection of the most statistically relevant features and linear kernel support vector machines for final layer classification. We compare DWNN with next-gen deep networks. Our dataset consists of 336 thermographic images, classified into healthy (no lesion), cyst, benign lesion and malignant lesion. Experimental results show that the 6-layer DWNN achieved accuracy, sensitivity, specificity, kappa and precision above 98%. These results show that DWNN are competitive deep architectures that can be used to optimise thermographic image analysis and clinical adoption.