Fig. 2

Proposed deep learning-based hierarchical diagnostic model (CLA-HDM) to non-invasively assess unexplained CLA. a Each sub-model takes BUS and CDFI images as inputs and assigns weights between different color channels in CDFI branch and pays attention to specific CDFI features under the guidance of BUS branch via attention mechanism. b For each test case, our model utilizes dual-modal ultrasound images as inputs each time, outputs hierarchical diagnostic task-related predictive probabilities and corresponding heatmaps to compare with and assist radiologists. CLA, cervical lymphadenopathy; BUS, B-mode ultrasound; CDFI, color Doppler flow imaging; AI, artificial intelligence