Özet:

In recent years, artificial intelligence studies, which have developed in parallel with computer hardware, have helped increase patient survival by preventing possible metastasis with early diagnosis of clinicians. There are many studies in literature that carry out diagnosis of cancer in the clinic. In these studies, machine learning and deep learning applications are frequently applied for cancer classification. Similarly, in this study, diagnosis of breast cancer over thermal images with the help of deep learning methods was discussed. The images used in the study were taken open access made available in the DMR-IR dataset. Some preprocessing was done on the images before manual and automatic segmentation methods were applied to segment the breast regions on images. In the manual segmentation process, the mask of the breast regions whose localization information was recorded with VIA was created and the segmentation was performed by subtracting the mask from the original image. In the automatic segmentation process, segmentation was done using Mask R-CNN and U-NET techniques. Segmentation performance analysis was performed for these two methods and classification operations were done with Mask R-CNN, which realized 0.9896 accuracy, 0.9413 Dice and 0.8900 Jaccard. Breast cancer classification was carried out using seven pre-trained architectures (InceptionV3, MobileNet, MobileNetV2, ResNet50, VGG16, VGG19 and Xception) with thermograms consisting of images segmented by manual and Mask-RCNN methods. As a result, MobileNet and InceptionV3 architectures provided the 100% classification success in test data accuracy, precision, sensitivity and F1 Score.

Anahtar Kelimeler: