Comparison and Ensemble Transfer Learning for Breast Cancer Classification in Ultrasound Images Using EfficientNet-B0, ResNet-50, and VGG-16

Abstract

Purpose: This study aims to compare three transfer learning architectures and develop an ensemble learning approach for breast cancer classification in ultrasound images. The objective of this study is to compare the three transfer learning architectures and evaluate an ensemble learning approach to determine the best model for distinguishing benign and malignant tumors.

Methods: This study used a combined dataset from the BUSI Kaggle challenge and clinical data from Bhakti Dharma Husada General Hospital in Surabaya, consisting of 1,632 images (1,048 benign and 584 malignant). Three pre-trained CNN models, EfficientNet-B0, ResNet-50, and VGG-16, with ImageNet weights, were evaluated using accuracy, precision, recall, and F1-score metrics. Preprocessing steps included resizing to 224×224, ImageNet normalization, and ROI segmentation, with ground-truth labels created using Roboflow while automatic segmentation was performed using DeepLabV3+. The ensemble model employed a soft voting method based on the average probability.

Result: The soft voting ensemble achieved an accuracy of 86.18%, precision of 85.21%, recall of 84.46%, and an F1-score of 84.40% (macro-averaged). Among the four evaluated models, ResNet-50 achieved the highest overall accuracy (87.40%).

Novelty: The novelty of this research lies in the complementary application of an ensemble of three transfer learning architectures (EfficientNet-B0, ResNet-50, and VGG-16) via the probability averaging soft voting method, enhanced by automatic segmentation preprocessing based on DeepLabV3+, and tested on the public BUSI dataset as well as local clinical data from Bhakti Dharma Husada General Hospital in Surabaya.