High accuracy hybrid CNN classifiers for breast cancer detection using mammogram and ultrasound datasets

Breast cancer is a significant cause of cancer fatality among women all over the world. Hence the detection of this disease at the initial stage works as a boon to the patient so that proper treatment can be provided. We have developed five new deep hybrid convolutional neural network-based breast cancer detection frameworks in this work. The proposed hybrid schemes exhibit better performance than the respective base classifiers keeping the combined benefits of both the networks. In addition, a probability-based weight factor ( w ) and threshold value ( β ) play a crucial role in making an efficient hybridization. Experimentally selected optimum threshold value ( β ) makes the system faster and more accurate. More importantly, unlike traditional deep learning methods, the proposed framework yields excellent performance even in small datasets. The proposed scheme is validated with datasets of two different breast cancer modalities: mini-DDSM (mammogram), BUSI and BUS2 (ultrasound). The experimental results demonstrate the superiority of the proposed ShuffleNet-ResNet scheme over the current state-of-the-art methods in all the mentioned datasets. Moreover, the proposed scheme achieves the accuracy of 99.17%, 98.00% for abnormality and malignancy detection in mini-DDSM respectively, and 96.52%, 93.18% for abnormality and malignancy detection BUSI datasets, respectively. BUS2 delivers 98.13% accuracy for malignancy detection. • Five new hybrid CNN frameworks have been proposed for breast cancer detection. • The combined benefits of classifier 1 and classifier 2 improve performance. • Emphasis is given to both abnormality and malignancy detection. • The efficacy of the system is validated using mammogram and ultrasound databases. • Moreover, the proposed ShuffleNet-ResNet framework gives the best performance.