An Insight-Driven Comparative Analysis of Deep Learning Models for Breast Ultrasound Image Classification
Early detection of breast tumors plays an important role in saving patients’ lives. The utilization of ultrasounds in medicine is common practice since they are noninvasive and allow for analysis of dense breast tissue, although it may be hard for a person to analyze ultrasound images manually due to their high level of noise. This work aims at designing the classification framework of breast ultrasound images using deep learning techniques that can automatically detect benign/malignant cases via four differential networks, which include Custom Convolutional Neural Network, VGG16, EfficientNetB0, and DenseNet201. A total number of 9016 images was used for training and testing purposes of the four models using numerous types of image preprocessing and augmentation techniques for improving accuracy. After comparing results of each architecture, it was shown that the VGG16 model gave the highest accuracy of 98%, while DenseNet201 and Custom CNN gave comparable results, however, in case of EfficientNetB0, lower accuracy was found among the other models. Overall, the experiments proved that the pre-trained model architectures provide more accurate results than other architectures and could significantly assist doctors in faster decision making.