Classification for Rectal CEUS Images Based on Combining Features by Transfer Learning

Abstract

It is very important to diagnose patients with rectal cancer, which can provide reference for the follow-up treatment. The gold standard for rectal cancer diagnosis is biopsy, but biopsy is invasive and risky. With the development of contrast-enhanced ultrasound (CEUS) technology, CEUS has become a reliable modality to diagnose rectal cancer. The degree of contrast enhancement can reflect the distribution of micro vessels inside the tumor. CEUS images are classified into three grades according to the inhomogeneity of enhancement inside rectal cancer. In this paper, we use deep learning and transfer learning to classify CEUS images. Features of rectal CEUS images were extracted by AlexNet, VGG16 and Resnet50. The extracted features were combined and normalized. A three-layer fully connected neural network was trained to classify the features of rectal CEUS images. The combination of features extracted by VGG16 and ResNet50 achieve 87.91% accuracy and AUC is 0.978.