2-Stage Convolutional Neural Network for Breast Cancer Detection from Ultrasound Images
Sagar Deep Deb, Arjun Abhishek, Rajib Kumar Jha · 2023
Breast Cancer is one of the most commonly occurring cancers in women. Detecting cancer at an earlier stage increases the chance of survival. Thus the development of an automatic Computer-Aided Diagnosis (CAD) system is essential for cancer diagnosis. Various researchers have used different image modalities for detecting cancer. This manuscript proposes a 2-stage Convolutional Neural Network (CNN) for detecting breast cancer from Ultrasound images. The CNN in the first stage is a customized DenseNet architecture whose primary task is to extract features from the input Ultrasound Images. The initial layers of the DenseNet network are pre-trained on the ImageNet dataset, while the final five layers are fine-tuned on the Breast Ultra Sound Image dataset. In the second stage, we used a bi-stream CNN with depth-wise separable layers to classify the features into three classes, namely, Normal, Benign, and Malignant. Conducting five-fold cross-validation using the proposed 2-stage CNN an accuracy of 85.38±1.21 was obtained. We have proved that the proposed network effectively classifies the input Ultrasound Images into three categories. Statistical significance test and Grad-CAM visualization is provided to prove the same.