Fig. 5: Deep neural network-enabled automatic biosensing.

a A CNN based on the VGG-16 architecture classifies the detection zone. It processes grayscale images of the asymmetric pattern to output the probability of a positive diagnosis (see Figure S9 for architecture). b, c A C-GAN automatically segments specific zones. The generator and discriminator are co-trained to optimize performance, with the generator producing segmentation maps resembling manually labeled images. d The network processes detection zone patterns and uses the C-GAN to segment the target area, removing artifacts and noise to facilitate concentration estimation. Three coffee-ring patterns from the same protein concentration are compared to demonstrate the method’s consistency (Figure S27). Network structures and training details are provided in Figures S10–S13. e Biosensor screening performance for N-Protein, PCT, CEA, and PSA protein shows standard LODs (red points) at 50 pg/mL, 100 pg/mL, 750 pg/mL, and 10 pg/mL, respectively. Probabilistic LODs (blue stars) are ~50 pg/mL, ~30 pg/mL, ~650 pg/mL, and ~3 pg/mL (see Supplementary Note for LOD definitions). Note, the blue dashed line shows the logistic regression fit. f An FC regression network predicts concentrations from crossline profiles by extracting features such as intensity, gradient, and coffee-ring patterns to establish a non-linear mapping between these features and concentration. g Predicted versus actual concentrations for four biomarkers (N-Protein, PCT, CEA, PSA) show excellent predictions by the FC network with minimal errors. Variations are due to testing uncertainties and protein degradation. Data are presented as mean ± SD, based on n = 3 crosslines. h The testing procedure involved spiking pooled human saliva diluted in PBS with N-Protein to create test samples. i Screening performance for N-Protein in saliva shows a standard LOD of 100 pg/mL and a probabilistic LOD of 50 pg/mL, outperforming equivalent LFIA tests by over two orders of magnitude (Figure S20). j Predicted versus actual concentrations for N-Protein mixed with human saliva closely match results obtained without saliva. All measurements were repeated at least three times. Concentration quantification was based on n = 5 random samples per concentration. Screening data analysis included n = 10 random samples.