Improving Generalization of ENAS-Based CNN Models for Breast Lesion Classification from Ultrasound Images

Neural Architecture Search (NAS) is one of the most recent developments in automating the design process for deep convolutional neural network (DCNN) architectures. NAS and later Efficient NAS (ENAS) based models have been adopted successfully for various applications including ultrasound image classification for breast lesions. Such a data driven approach leads to creation of DCNN models that are more applicable to the data set at hand but with a risk for model overfitting. In this paper, we first investigate the extent of the ENAS model generalization error problem by using different test data sets of ultrasound images of breast lesions. We have observed a significant reduction of overall average accuracy by nearly 10% and even more severe reduction of specificity rate by more than 20%, indicating that model generalization error is a serious issue with ENAS models for breast lesion classification in ultrasound images. To overcome the generalization error, we examined the effectiveness of a range of techniques including reducing model complexity, use of data augmentation, and use of unbalanced training sets. Experimental results show that different methods for the tuned ENAS models achieved different levels of accuracy when they are tested on internal and two external test data sets. The paper demonstrates that ENAS models trained on an unbalanced dataset with more benign cases tend to generalize well on unseen images achieving average accuracies of 85.8%, 82.7%, and 88.1% respectively for the internal and the two external test data sets not only on specificity alone, but also sensitivity. In particular, the generalization of the refined models across internal and external test data is maintained.