A variance-reduction method for thyroid nodule boundary detection on ultrasound images

To perform computer-aided diagnosis of thyroid nodules on ultrasound images, the nodule´s location and its margin should be clearly defined. However, due to the nodule´s biological characteristics, echo structure and quality, operator´s subjective factors and operating conditions, identification of thyroid nodule boundary becomes quite difficult. In addition, manual identification of nodule boundary heavily relies on physician´s subjective judgment. Even the same physician could give different results on the same image at different times. In this study, we proposed a novel and automatic method for thyroid nodule boundary detection based on Variance-Reduction statistics. Based the operator´s initial inputs of the nodule´s major and minor axes, the region of interest (ROI) is first generated. With grayscale values of pixels in the ROI, we then implement an algorithm to automatically detect the nodule boundary. The proposed method is validated with ultrasound images of 433 thyroid nodules, and the effectiveness of the method is shown by comparing the two boundary error metrics, the Hausdorff distance (HD) and the mean absolute distance (MD), to previously published results.