combination of transfer and residual learning: effective tool to improve the performance of deep neural networks in breast cancer detection in thermograms

Recently thermography imaging has been introduced as an effective tool for early detection of breast cancer. a vital step of this process is classifying captured images into healthy and sick categories which its main challenge is the lack of visual difference between these two types of images in many samples. although deep learning methods may lead to better results than their classical alternatives in classifying thermography images but unfortunately two factors may hamper their effectiveness including the gradient vanishing as well as high dependence of the performance due to the volume of training and test data. in this article, the combination of residual and transfer learning schemes is utilized in order to improve the performance of deep learning schemes against two above problems. the obtained results from applying the proposed scheme on real thermography images show that it may improve the detection parameters at least 13.5, 7.7 and 18.7 percent superiority against the best non-transfer-learning results in terms of accuracy, sensitivity and specificity respectively.