| Abstrakt: |
Computer-Aided Diagnosis (CAD) system is preferred for automatic thyroid tumor ultrasound image characterization instead of manual assessment by the experts. Segmentation and control despeckling are the important pre-processing stages required to develop an effective CAD system. This work primarily aims to design an efficient CAD system for thyroid tumor characterization using ultrasound (US) images. Here, Edge Preserving Smoothing despeckling filter and encoder decoder-based ResNet50 segmentation model are used as pre-processing stages of the proposed CAD system to enhance its performance for thyroid tumor characterization. Extracting the image features using pre-trained models effectively captures the underlying textural and morphological characteristics exhibited by thyroid tumors in ultrasound images. The pre-trained- models learn by automatic feature extraction representing the underlying characteristics using multiple stages by convolution with various filters. The pre-train-based neural network classifies tumors more accurately due to learning the extract multiple sets of features. Accordingly, fifteen (15) deep learning-based pre-trained models have been utilized in the present work to extract information from the thyroid tumor US images and train the PCA-SVM classifier. These pre-train models have been taken from different categories of deep learning algorithms, including Series / DAG / Lightweight architectures, namely AlexNet, VGG16, VGG19, Darknet19, Darknet53, GoogleNet, DenseNet201, ResNet18, ResNet50, ResNet101, EfficientNetb0, NasNetMobile, MobileNet, SqueezeNet, and ShuffleNet for characterization of thyroid tissues. An exhaustive set of experiments have been conducted, and the best-performing pre-trained models have been selected as optimal feature extractors based on classification accuracy. Thus, the features extracted from the best-performing pre-trained network, i.e., ResNet50, are fed to the PCA-SVM classifier to yield an efficient CAD system for classifying TTUS images. The optimal CAD design proposed in the present work yields 99.5% classification accuracy to distinguish between benign and malignant thyroid tumors. [ABSTRACT FROM AUTHOR] |
Full text from SpringerLink