Fig. 5: Neural network detects and classifies glomeruli using only Stimulated Raman Histology. | Nature Communications

Fig. 5: Neural network detects and classifies glomeruli using only Stimulated Raman Histology.

From: Label-free multimodal optical biopsy reveals biomolecular and morphological features of diabetic kidney tissue in 2D and 3D

Fig. 5

a Adjacent kidney sections from the control group were imaged using label-free SRH and ground-truth histological staining for both H&E and PAS. Images were imported into the HALO AI platform by IndicaLabs and single glomeruli were manually annotated in each image. In all cases, SRH efficiently detected glomeruli as well as histologically stained samples. b 16 representative images of glomeruli were cropped from various patient samples in both control and diabetic groups. c The network architecture of DenseNet201 is briefly explained using the MATLAB R2024a release implementation. d The DenseNet201 network was trained using 5 Epochs, with 25 iterations per Epoch using a single Nvidia RTX 4080 and took less than 15 min. e The cross validation matrix shows an accuracy of 0.9762, sensitivity of 1, specificity of 0.9583, and precision of 0.9474, resulting in an F1 score of 97.22%. Images were bootstrapped from augmented initial image sets by transposition, increased contrast, added white speckled noise, and gaussian blurring. Half of the images were randomly selected for testing purposes. Scale bar: 250 μm. H&E Hematoxylin and Eosin, PAS Periodic Acid Schiff, SRH Stimulated Raman Histology.

Back to article page