Deep Learning Based Classification Model for Thyroid Malignancy Detection in Ultrasonic Images using Convolution Neural Network

Abstract

This work is a focus in the recognition of thyroid cancer by using medical images from ultrasound imaging modality. In ultrasonic images, those malignant (cancer)  parts of the body are normally blurred or asymmetrical in figure.  The feature of tumor (cancer) area is very parallel to benign or malignant tissues. Ultrasonic image is quite essential in the imaging techniques for finding the thyroid diagnosis. Thyroid cancer is the 2nd largest disease in the endocrinology field. We are developing a classification model that will be more suitable for detecting papillary thyroid cancer by ultrasound with improved accuracy. For experimental setup we are running with Python 3.6.5 for deep learning model development and image processing on 6th generation i7 processor with 32 GB RAM and 2TB SSD hardware. In this endeavor we are working with 9160 images with split and validation in 7328 by 1832 ratio from DDTI dataset provided by Universidad Nacional de Colombia , CIM@LAB and IDIME ( Instituto de Diagnostico Medico).

Keywords: Deep learning, ultrasonic image, Convolutional neural network (CNN), feature fusion, Thyroid cancer.