A deep learning approach for three-dimensional thyroid nodule detection from ultrasound images

Currently, thyroid diseases are prevalent worldwide; therefore, it is necessary to develop techniques that help doctors improve their diagnostic skills for such diseases. In previous studies, 2-dimensional convolutional neural network (2D CNN) techniques were employed to classify thyroid nodules as benign and malignant without detecting the presence of thyroid nodules in the obtained ultrasound images. To address this issue, we propose a 3-dimensional convolutional neural network (3D CNN) for thyroid nodule detection. The proposed CNN exploits the 3D information and spatial features contained in ultrasound images and generates distinctive features during its training using multiple samples, even for small nodules. In contrast, a 2D CNN only depends on spatial features. In this study, we used two datasets of 2210 ultrasound images obtained from the Sultan Abdul Aziz Shah Hospital in Malaysia, and a public open dataset, Digital Database Thyroid Image (DDTI). We created folders containing three images each, processed the images and extracted volumetric features suitable for the 3-dimensional convolutional neural network (3D CNN). The proposed model was assessed using four metrics: accuracy, recall, precision and F1 score. The results showed that the accuracy of the model in predicting the presence of thyroid nodules in ultrasound images was 96%. In conclusion, this study could help radiologists in hospitals and medical centres in classifying ultrasound images and detecting thyroid nodules.

Text 125487.pdf - Published Version Available under License Creative Commons Attribution. Download (36MB) Official URL or Download Paper: https://www.techscience.com/CMES/v146n3/66786

Actions (login required)

View Item