Computer Aided Diagnosis System for Classification of Abnormalities in Thyroid Nodules Ultrasound Images using Deep Learning

摘要

Thyroid nodules, which are abnormal cell growth that forms a lump in the thyroid gland occurs in more than 50% of the adult population. While ultrasonography is a commonly used for diagnosis because it is non-invasive, non-radioactive, relatively inexpensive and widely available, the usual visual interpretation results in subjective interpretation, inter-observer variability and it's a time-consuming process. In addition, different stages of malignancy may not be detected, even in existing Computer Aided Diagnosis (CAD) systems. Therefore, there is need for CAD systems that can classify thyroid nodules into multiple stages of malignancies. In this work, we developed a classification system that classifies ultrasonic thyroid images into Thyroid Imaging Reporting and Data System (TIRADS) classes using convolutional neural network. Histogram equalization and adaptive filtering were used to improve the image contrast and noise removal respectively, while transfer learning with Alexnet was used to classify into 6 TIRADS classes. The developed system achieved an overall performance of 92% accuracy, 91% sensitivity and 99% specificity. The work will provide a second opinion, aid early and more accurate detection, thereby enhance treatment and management procedures.