Fig. 3: Assessment of the performance of MD-nets in quantitatively evaluating cell morphology using human sperm cells as a clinical model.

a, Linear regression plot (n = 10). The dotted line represents the line of identity. The solid red line represents the best-fitting straight line on the available data points identified using a least-squares analysis. The dotted red line represents the 95% confidence interval of the fitted line. The equation represents the line equation of the fitted line and r represents the Pearson’s correlation coefficient (P = 0.004). b, The absolute difference between morphological scores of human semen samples (SD4) when measured using automated MD-net and manual-based assessment (national average) performed using conventional manual microscopy. n = 10. c, Comparison of distributions of SIFT feature points per image identified across the various domain-shifted datasets of microscopic sperm images. SD4, n = 86,440; SD3, n = 18,972; SD2, n = 19,668; and SD1, n = 20,554. Feature estimates are not separated by classes as manual annotations are unavailable. The dashed lines represent the median and the dotted lines represent quartiles. d–f, Bland–Altman tests comparing morphology scores estimated on the basis of manual microscopy analysis and MD-net using sperm samples imaged using a benchtop microscope (d) (SD3, n = 40); a portable 3D-printed microscope (e) (SD2, n = 48); and a portable smartphone-based microscope (f) (SD1, n = 47). The dotted lines indicate the mean bias and the blue region within the blue dashed lines indicates the 95% limits of agreement. g, t-SNE plots illustrating source and target clustering achieved by MD-net for the four different sperm datasets. h, Examples of sperm images that were collected using the different optical instruments along with the associated saliency maps obtained from the MD-net feature extractor.