Breast Cancer Detection Using Various Classification Models Combined with Transfer Learning

Abstract

When mortality rates are considered, breast cancer is among the most common forms of cancer worldwide. As with every cancer, early diagnosis and treatment are the most effective method of preventing breast cancer. Artificial intelligence’s development in the health sector has decreased the margin of error compared to the old mammographic and manual methods. It has now begun to be obtained much earlier and with a low margin of error. Several researchers have worked on the segmentation and categorization of breast cancer using various imaging modalities. One of the imaging methods with the highest sensitivity for diagnosis is the ultrasonic imaging modality. For this purpose, ultrasound images are used in this study, and there are three categories of images in the dataset: normal, benign, and malignant images. This study aims to develop a technique for spotting and diagnosing breast cancers using ultrasound images. Deep learning techniques are key alternatives to feature-based approaches for overcoming their numerous drawbacks. Convolutional neural network models that have been previously trained and machine learning classifiers are used together in this study. This research compares six distinct pre-trained models: MobileNetV1, MobileNetV2, DenseNet121, DenseNet169, ResNet50, and ResNet101, and various classifiers such as Support Vector Machine, Adaptive Boosting, K-Nearest Neighbors, Random Forest, Bootstrap Aggregating and Extreme Gradient Boosting. As a result of the experiments, it was seen that the highest accuracy scores are achieved by using the MobileNetV2 pre-trained model when looking at overall accuracy percentages with nine different classifiers. In addition, when these nine different classifier algorithms are examined among themselves in particular MobileNetV2, Support Vector Machine, K-Nearest Neighbor and Long Short-Term Memory gave the best accuracy results.