Breast Cancer Detection Using B-Mode and Ultrasound Strain Imaging

Deep learning (DL) methods have recently been used more and more in US breast imaging. Breast Imaging-Reporting and Data System (BI- RADS) score is used to categorize B-mode pictures. DL algorithms identify lesions’ patterns, such as their direction, and border. The convolutional neural network (CNN), one of several DL neural networks, is most frequently applied to image categorization. However, gathering a sizeable breast ultrasound (US) dataset is time-consuming and occasionally impossible since training a CNN from scratch requires a big, labeled image collection. Here, this research introduces a new method for classifying benign and malignant lesions in breast US pictures utilizing ensemble TL using a mix of B-mode and SE images. In order to extract features for a more accurate diagnosis, this research attempts to stack-wise merge B-mode and SE images (referred to as B-SE for concision). For excellent prediction performance, two classification models are integrated into one classifier in this instance. The Kaggle, ImageNet (a huge dataset of annotated natural photographs) dataset are used to train the CNN model, while the ImageNet (a small collection of annotated breast US images) dataset is used to fine-tune (optimize) the model. AlexNet and deep residual network (ResNet) are two CNN models that are frequently employed because they are well known, have low false-positive rates, and are accurate at classifying medical images. The highest accurate model will be deployed to web applications where doctors and patients can easily get predictions on their input.