Ensembles of deep learning networks for classification of breast ultrasound images

Abstract

This paper examines the problem of three-class classification of breast ultrasound images (benign, malignant, and cystic). A comparative analysis of six deep learning models (AlexNet, ResNet-18, InceptionResNetV2, ViT, Swin, and ConViT) and their two-component combinations is presented, using four machine learning classifiers (KNN, SVM Linear, SVM Quadratic, and Random Forest). It was found that the best result among individual models was achieved using the Inception v2 model. Among the four classifiers studied, SVM Quadratic and Random Forest demonstrated the best performance for most feature combinations. It was shown that using a feature ensemble can significantly improve results for architectures with inherently lower accuracy: The highest sensitivity to malignant lesions was observed for the InceptionResNetV2 + Swin Transformer combination with the Quadratic SVM classifier (91.46%). It is demonstrated that the proposed approach, based on combining features from different architectures, enables high-accuracy three-class classification of breast ultrasound images.