Abstract:
Deep Convolutional Neural Networks (Deep-CNNs) algorithms have demonstrated superior performance over different machine learning approaches in the early-stage detection of breast cancer using mammography images in recent years. These algorithms have aided radiologists in detecting suspicious breast masses and other key characteristics in mammography images with greater accuracy. However, the accuracy of such deep learning models in detecting breast cancer highly relies on a number of factors like the quantity and quality of mammography images, the methodology used to extract abstract features, the selection of network architecture, and the appropriate hyperparameters for a chosen deep learning strategy. Incorrect selection of such factors might lead to unsatisfactory results in breast cancer diagnosis and prediction. The ensemble of CNN architectures has been successful in improvizing model performance designed for complex disease analysis. Our paper aims to create an ensemble Deep CNN-based mammography image classification model by concatenating the features extracted by VGG16, VGG19 and ResNet-50 pre-trained CNNs. The ensemble features are classified with ensemble machine learning classifier frame (Support Vector Machine, K-Nearest Neighbor, Decision Tree, Random Forest, Gaussian Naive Bayes) to classify the mammography images. The results obtained with our proposed model are compared with existing DL-based breast cancer detection methods using two publicly accessible mammography datasets [Mammographic Image Analysis Society (MIAS), Digital Database for Screening Mammography (DDSM)]. The blend of ensemble CNNs as feature extractors and ensemble machine learning classifiers resulted in classifications of breast masses into binary as well as multi-class with better accuracy. The proposed model outperformed some of the existing models by achieving 96% accuracy with DDSM and 99% accuracy for MIAS datasets. The paper also describes the workflow of deploying the trained ensemble DL mammography image classification model on a Raspberry Pi 3B embedded device. The deployment of trained DL models on low-cost, low-computational embedded devices opens a wide room for researchers to develop faster and cost-effective Computer-Aided Detection systems for disease diagnosis
