Fig. 1: Start-to-end deep learning-guided optical trapping workflow for single-EV analysis. | npj Biosensing

Fig. 1: Start-to-end deep learning-guided optical trapping workflow for single-EV analysis.

From: Accurate label free classification of cancerous extracellular vesicles using nanoaperture optical tweezers and deep learning

Fig. 1

a Exosomes bud inward into MVBs and are released when MVBs fuse with the plasma membrane. (LE: intraluminal vesicles; MVBs: multivesicular bodies; EE: early endosome; TGN: trans-Golgi network). EV samples are diluted, placed on the sample stage, and covered by the double nanohole (DNH) aperture. Single EVs are trapped using a DNH optical tweezer, with trapping signals detected by a photodiode. b Representative trapping signals, their probability density functions, and power spectral density boxcar averages are shown, each fitted with a Lorentzian curve across different EV lines. c Signals are pre-processed and split (8:1:1) into training, validation, and test sets before being encoded and input into deep learning networks. The model is trained to optimize classification performance and extract latent features for visualization and downstream analysis. d Final outputs include EV class predictions and visualizations of the latent feature space.

Back to article page