TY - JOUR
T1 - PAIM (πM)
T2 - portable AI-enhanced fluorescence microscope for real-time target detection
AU - Jiao, Ziao
AU - Zang, Zhenya
AU - Wang, Quan
AU - Chen, Yu
AU - Xiao, Dong
AU - Li, David Day Uei
PY - 2023/8/31
Y1 - 2023/8/31
N2 - We proposed a portable AI fluorescence microscope (πM) based on a webcam and the NVIDIA Jetson Nano (NJN), integrating edge computing techniques for real-time target detection. πM achieves a physical magnification of ×5 and can resolve 228.1 lp/mm USAF features. Prepared microscopic samples and fluorescent polystyrene (PS) beads can be imaged clearly. πM’s body was fabricated by a 3D printer, weighing ~250 grams with dimensions of 145mm × 172 mm × 144 mm (L×W×H), costing ~$300. It has a similar brightfield imaging quality compared to benchtop microscopes (~$13,000). The customized convolution neural network (CNN) inside the NJN can realize feature extraction, real-time PS bead counting, and red blood cell counting without data transfer and offline image processing. Compared with two model-free image processing methods (OpenCV and CLIJ2), our CNN method is robust in bead counting at different concentrations. Six aggregated beads can be correctly counted with 80% accuracy. Regarding feature extraction and human RBC counting, our CNN also obtained closer results to the ground truth (GT) than the CLIJ2 method (GT: 201; CNN: 196; CLIJ2: 189). With a miniature size and real-time analysis, πM has potential in point of-care testing, field microorganism detection, and clinical diagnosis in resource-limited areas.
AB - We proposed a portable AI fluorescence microscope (πM) based on a webcam and the NVIDIA Jetson Nano (NJN), integrating edge computing techniques for real-time target detection. πM achieves a physical magnification of ×5 and can resolve 228.1 lp/mm USAF features. Prepared microscopic samples and fluorescent polystyrene (PS) beads can be imaged clearly. πM’s body was fabricated by a 3D printer, weighing ~250 grams with dimensions of 145mm × 172 mm × 144 mm (L×W×H), costing ~$300. It has a similar brightfield imaging quality compared to benchtop microscopes (~$13,000). The customized convolution neural network (CNN) inside the NJN can realize feature extraction, real-time PS bead counting, and red blood cell counting without data transfer and offline image processing. Compared with two model-free image processing methods (OpenCV and CLIJ2), our CNN method is robust in bead counting at different concentrations. Six aggregated beads can be correctly counted with 80% accuracy. Regarding feature extraction and human RBC counting, our CNN also obtained closer results to the ground truth (GT) than the CLIJ2 method (GT: 201; CNN: 196; CLIJ2: 189). With a miniature size and real-time analysis, πM has potential in point of-care testing, field microorganism detection, and clinical diagnosis in resource-limited areas.
KW - PAIM (πM)
KW - portable AI-enhanced fluorescence microscope
KW - real-time target detection
UR - https://www.sciencedirect.com/journal/optics-and-laser-technology
U2 - 10.1016/j.optlastec.2023.109356
DO - 10.1016/j.optlastec.2023.109356
M3 - Article
SN - 0030-3992
VL - 163
JO - Optics and Laser Technology
JF - Optics and Laser Technology
M1 - 109356
ER -