12 Detection of breast cancer using deep neural networks with transfer learning on histopathological images
Kerim Kürşat Çevi̇k, Emre Dandıl, Süleyman Uzun, Mehmet Süleyman Yıldırım, Ali Osman Selvı · 2021
Breast cancer remains as the second leading cause of cancer death in women worldwide. The basal-like carcinoma is defined as an important breast cancer subtype and associated with its protein and gene expression profile. There are five main breast subtypes, namely, normal breast-like subtype, luminal subtype A, HER-2 overexpression subtype, luminal subtype B and basal-like subtype. Breast cancer includes five molecularly defined subclasses, one of which is basal-like tumors. In recent years, basal-like tumors have taken great attention of researchers due to limited therapy opportunities they offer. This tumor subtype tends to occur more frequently in younger (<50) patients and almost 15% of all breast cancers are caused by this tumor subtype. Basal-like tumors are expressed by their distinctive immunophenotypic, morphologic, clinical and genetic features. However, no widely accepted consensus could be reached on definition and common clinical identification of the subtype of breast cancer. In addition, there is no widely accepted way on how to systematically classify this complex group of tumors. As a cost-effective approach, immunohistochemical markers have been used to classify the basal-like breast cancer (BLBC). Unfortunately, these markers only have a 60% accuracy. In this study, in order to improve the accuracy of BLBC detection on histopathological images, the results of performance tests using a deep-learning-based approach on multiple GPUs are presented. Experimental studies are performed with the help of five different pretrained deep neural networks such as ResNet18, ResNet50, AlexNet, SqueezeNet and InceptionV3 using global histopathological image dataset breast histopathology images. Parallel computing performances are also evaluated using GPUs during the experimental tests. As a result, 70.18%, 95.13%, 77.55%, 75.89% and 96.94% accuracy rates are achieved for each pretrained network, respectively.