Thyroid　tumor　is　a　common　clinical　disease.　Fully　automated　computer-aided　diagnosis　of　thyroid　nodules　has　important　clinical　significance.　However,　most　of　the　previous　works　focused　on　the　classification　of　nodules,　benign　or　malignant,　but　how　to　locate　the　nodules　in　the　ultrasound　images　is　rarely　studied.　This　paper　focuses　on　the　problem　of　thyroid　nodule　detection,　aiming　to　achieve　a　fully　automated　method　for　delineating　the　nodule　bounding　box　from　the　thyroid　ultrasound　image.　We　propose　a　nodule　detection　algorithm　based　on　the　convolutional　neural　network　for　this　task.　We　explore　the　performance　of　nodule　detection　in　three　aspects:　multi-scale　prediction　architecture　design,　loss　function　design　and　postprocessing　method.　This　method　is　evaluated　on　clinical　data　and　compared　to　the　ground　truth　labeled　by　doctors.　The　experimental　results　show　that　this　proposed　method　can　achieve　88.08　%　AP　with　90.08　%　overall　recall.　©　2019　IEEE.