Utilizing Transfer Learning-Based Algorithms for Breast Ultrasound Data in Multi-Instance Classification

In the last decade, researchers have placed substantial emphasis on automating the analysis and detection as well as classification of the breast masses in an early state. While ultrasound imaging has gained prominence as a diagnostic tool for breast cancer, its predictive accuracy still relies on the expertise of specialists. This paper seeks to contribute to the academic body of knowledge regarding the classification of breast masses within a multi-instance framework. Additionally, it aims to shed light on potential directions for future research in this field.The dataset utilized in this research encompasses 780 breast ultrasound images, including 437 with benign masses, 218 with malignant masses, and 133 depicting normal tissue characteristics. Various pre-trained architectures were implemented, including VGG16, InceptionV3, ResNet50, MobileNetV2, InceptionRes-NetV2, and DenseNet121. The top-performing model turned out to be MobileNetV2, achieving an accuracy rate of 84% and an "one vs rest"AUC score of 0.946. This result represents a 1% improvement in the current state of the art and underscores the significance of the AUC "one vs rest"score in the context of multi-instance breast mass classification.