Deep Learning for Analyzing Thyroid Nodule Malignancy Based on the Composition Characteristic of the Ultrasonography Images

Abstract

There are ten characteristics of ultrasound thyroid nodules, and each one has its categories. One particular characteristic is composition characteristic, as the occurrence of its categories can be one of the lead to the incidence of malignancy. This study developed a method to help experts identifying the categories of the composition ultrasound characteristic using data collected from the Department of Radiology RSUP Dr. Sardjito. This dataset was already cropped by the experts, leaving only the thyroid nodule area as the region of interest. The dataset of ultrasound images was going to pre-processing first to remove the labels, markers, and unnecessary artifacts. To further remove any unnecessary artifacts, the pre-processed image was segmented. Afterward, the data augmentation begins using the synthetic minority over-sampling technique (SMOTE). The augmentation result was sent to LeNet to be classified into three categories those are cystic, solid, and complex. The testing result outperformed previous studies with 92% accuracy, 85.71% sensitivity, 92.50% specfficity, 93.5S% PPV, 95.09% NPV, 0.918 F Score, and 146. 4s testing time.

Ähnliche Arbeiten