Fig. 2. Results of the simulation study and in vivo chicken embryo CAM imaging study.

a Simulation results of conventional localization, Deep-ULM¹⁴, mSPCN²¹, and LOCA-ULM with low ($0.16\mathrm{MBs}/{\lambda }^2$) and high ($0.35\mathrm{MBs}/{\lambda }^2$) MB concentrations. Ground truth MB positions are marked by red circles, conventional localization by yellow $\times$, Deep-ULM by green $\times$, mSPCN by blue $\times$, and LOCA-ULM by cyan $\times$. b Comparison between conventional localization, Deep-ULM, mSPCN, and LOCA-ULM was performed on simulated frames at increasing MB concentrations, using three performance metrics: MB detection accuracy, MB missing rate, and localization error. The error bars indicate the standard deviation of inference over 500 frames. c–e Comparisons among conventional localization, LOCA-ULM^Experimental, and LOCA-ULM^LSGAN MB localization in in vivo CAM imaging. c Optical microscopy image of the CAM surface microvessel, along with MB localization images reconstructed by conventional localization and LOCA-ULM^LSGAN. n = 1 experiment. d The ROI selected from the optical image and the corresponding ground truth vessel segmentation. Magnified view of the MB localization images marked by the white ROI for conventional localization, LOCA-ULM^Experimental, and LOCA-ULM^LSGAN. e The vessel filling (VF) a percentage of conventional localization, LOCA-ULM^Experimental, and LOCA-ULM^LSGAN, as a function of the number of frames (Methods). f Comparison of computational times for MB localization across varying FOV sizes using conventional ULM, Deep-ULM, mSPCN, and LOCA-ULM. Computational times encompass the MB localization process, which includes normalized cross-correlation and regional maximum search for conventional ULM, a network forward pass for LOCA-ULM, and a network forward pass with centroid calculation for both Deep-ULM and mSPCN. Source data are provided as a Source Data file.