An Improved Ultrasound Tumor Segmentation Using CNN Activation Map Clustering and Active Contours
Revathy Sivanandan, J. Jayakumari · 2020
Ultrasound (US) imaging has become a primary tool for diagnosis of tumors mainly owing to its non-ionizing procedure but effective analysis is challenging due to the presence of inhomogeneties, speckle, low contrast and complex echogenic patterns. Manually defined predicates for feature extraction will not suffice, where deep learning can play a pivotal role. In this work, Convolutional Neural Networks (CNN) have been used for US breast tumor feature extraction and classification as benign/ malignant. To make the network robust to noise variations in the image and as an improvement over existing methodologies, neutrosophic preprocessing and enhancement were performed and the enhanced images were appended to the dataset and the original image during training. Using neutrosophic preprocessing alone had increased the validation accuracy of the CNN from 0.84 to 0.92, while using neutrosophic preprocessing and augmentation had increased the accuracy to 0.99. GoogLeNet with pretrained parameters were chosen as the preferred architecture for the work, with the predictions being validated using class activation maps that showed that the CNN was able to learn the features related to the tumor accurately. The final layers' activation maps are clustered using fuzzy c-means clustering, which act as an initial contour for active contour segmentation using localized energies. The clustered outputs also qualify as a validation method for the prediction accuracy of the CNN. Using the proposed segmentation, time and computational complexity can be reduced which otherwise is high for traditional encoding cum decoding based CNN segmentation architectures.