A Neural Network for Thyroid Segmentation and Volume Estimation in CT Images

Abstract

Thyroid region segmentation and volume estimation is a prerequisite step to diagnosing the pathology of the thyroid gland. In this study, a progressive learning vector quantization neural network (PLVQNN) combined with a preprocessing procedure is proposed for automatic thyroid segmentation and volume estimation using computerized tomography (CT) images. The preprocessing procedure is used to extract the region of interest (ROI) of thyroid glands and exclude non-thyroid glands based on thyroid anatomy. The PLVQNN contains several learning vector quantization neural networks (LVQNNs), each responsible for segmenting one slice of a thyroid CT image. The training of the PLVQNN is conducted starting from the LVQNN of most reliable (middle) slices, where the thyroid has the largest region. The training then propagates upwards and downwards to adjacent LVQNNs using the results of the middle slice as the initialization values and constraints. Experimental results show that the proposed method can effectively segment thyroid glands and estimate thyroid volume.