Determination of malinity by deep learning in some breast lesions through active thermograms

Artificial intelligence applications, which have increased in parallel with advances in computer hardware, are frequently preferred in cancer detection in biomedical field because they give rapid and high accuracy results, enable early diagnosis and prevent possible metastasis. In this thesis, segmentation and classification of breast regions were studied by thermal breast images. The images taken from an online data set were preprocessed before they were given to deep learning models. Then, the mammary regions were segmented and clipped from the outer areas for further training procedures. After this, two different methods were tried for classification. In the first method, the breast regions were manually segmented and classified by applying transfer learning. In the second method, lesion classification was studied by applying transfer learning after automatic segmentation of breast regions. For the automatic segmentation process used in the second method, U-Net and Mask R-CNN techniques were tried and the study continued with Mask R-CNN method because of higher performance. For the training of both methods used in the study, transfer learning was applied with the pre-trained InceptionV3, MobileNet, MobileNetV2, ResNet50, VGG16, VGG19 and Xception architectures. When the findings of the study were examined, the Xception architecture (100%) gave the highest training accuracy, precision and sensitivity. When the test performances of the trained architectures were examined, 100% accuracy, highest precision and sensitivity were obtained by InceptionV3 and MobileNet architectures. Among the architectures used, Xception and VGG architectures were the architectures that needed the longest training time, while the architecture that gave the fastest test results was the MobileNet architecture.