Abstract:

Breast cancer can progress silently in its early stages and frequently without noticeable symptoms. However, it poses a serious risk to women. It is imperative to recognize this potential health concern to mitigate it early. In the last few years, Convolutional Neural Networks (CNNs) have advanced significantly in their ability to classify images of breast cancer. Their capacity to automatically extract discriminant features from images has enhanced the performances and accuracy of image classification tasks. They outperform state-of-the-art techniques in this area. Furthermore, complicated models that were first learned for certain tasks can be easily adapted to complete new tasks by using transfer-learning approaches. However, deep learning-based categorization techniques could experience overfitting issues, particularly in cases where the dataset is small. The primary goal of this work is to investigate the performances of certain deep learning models to classify breast cancer images and to study the effects of data augmentation techniques, such as image rotation or displacement when utilizing a transfer learning approach. Using certain image datasets, the ResNet18, Resnet50, and VGG16 models demonstrated accuracy improvements, according to our experimental results.

Keywords: