Transfer Learning and Handcrafted Features Ensembles for Ultrasound Breast Cancer Image Classification

Abstract

Breast cancer is the most commonly diagnosed cancer in women. Its diagnosis via ultrasound imaging largely depends on the technical skill of the radiologist. This study developed a binary classification system for breast lesions, combining transfer learning models and handcrafted features in ultrasound images. Pre-trained CNNs like InceptionV3, EfficientNetB4, ResNet50, and VGG16 were used, along with SVM-classified handcrafted features. Models were individually analyzed and combined using late-fusion ensembles. ResNet50 achieved an F1-score of 81.97%. The best late-fusion ensemble model reached an F1-score of 83.90%. In the cross-dataset evaluation, the top late-fusion ensemble model in the development dataset scored an F1-score of 88.70% and 78.20% in the test BUSI and BUID datasets, respectively. These results emphasize the robust potential of a late-fusion ensemble that combines CNN transfer learning and handcrafted features to classify breast lesions in ultrasound images.