Thyroid carcinoma is one of the most common endocrine diseases globally, and the incidence has been on the rise in recent years. Ultrasound imaging is the primary clinical method for early thyroid nodule diagnosis. Regions of interest (ROIs) of nodules in ultrasound images are difficult to detect because of their irregular shape nand vague margins. Accurate real-time thyroid nodule detection can provide ROIs for subsequent nodule diagnosis automatically, avoid variabilities between the subjective interpretations and inter-observer effectively and alleviate the workloads of medical practitioners. The aim of this study was to present a reliable, real-time detection method based on the Faster R-CNN (region-based convolutional network) framework for accurate and fast detection of thyroid nodules in ultrasound images. Our study proposed a faster and more accurate thyroid nodule detection method based on the Faster R-CNN framework by adding three strategies: feature pyramid, spatial remapping and anchor-box redesign. Specifically, the network takes raw ultrasound images as inputs and generates boxes with positions and the possibilities that these boxes contain thyroid nodules. The proposed method could locate and detect target nodules accurately with a mean average precision of 92.79% with more than 9000 patient images. In addition, the detection rate has accelerated to >16 frames per second, four times faster than that of the initial network. Therefore, it can meet the requirements of clinical application. The performance of the fivefold cross-validation was also accurate and robust. The proposed automatic thyroid nodule detection method yields better performance in accuracy and detection speed, which indicates the potential value of our method in assisting clinical ultrasound image interpretation.