Breast Cancer Detection on Histopathology Images Using Pre-trained Computer Vision Models

Breast cancer is the most common type of cancer worldwide. According to the World Health Organization (WHO), there were 7.8 million women alive in 2020 who had been diagnosed with breast cancer, and it has claimed more women's lives than any other kind of cancer. With the recent rise of artificial intelligence, breast cancer detection using deep learning techniques is getting more popular. However, creating a deep learning model for a specific task from scratch costs a lot of time and money. Transfer learning is a well-known method that can make deep learning developments more efficient by leveraging pre-trained models. Using the BreakHis dataset, this paper will compare three cutting-edge pre-trained computer vision models: DenseNet, RegNet, and BiT, in predicting malignant or benign tumor tissue from breast histopathology images to determine which model is better for that specific task. Although the DenseNet model achieves the highest score with 93.7% Area Under the ROC Curve (AUC) and 97.4% Average Precision Score (APS), the BiT model is more suitable for deployment in a real-world setting since it can predict more malignant cases correctly than the other two models with a sensitivity score of 90.79%.