Label-free imaging and classification of live P. falciparum enables high performance parasitemia quantification without fixation or staining
Fig 6
Wavelength- and slice-dependent classification statistics on the custom UV microscope.
Top: Overall accuracy (left), FPR for rings (center), and recall for rings (right) were selected for detailed breakdown as a function of wavelength (columns), and focal offset (rows). The top row in each table shows the result of processing the data with our maximum confidence strategy for focal plane selection on a cell-by-cell basis. Bottom: The same statistics were plotted for a single focal plane on the commercial light microscope at 40× magnification. For each grouping by microscope and statistic, cells are color-coded on a scale from blue (lowest performance) to yellow (highest performance). All statistics were derived from multi-wavelength datasets, which consisted of distinct but partially-overlapping sets of data with single-wavelength datasets (see Table 1). In both cases, each wavelength-specific classifier was trained and validated uniquely on images at its particular wavelength. To ensure a fair comparison between the wavelengths, all three wavelengths for each microscope shared the exact same random partitions between training and validation, ie. the same distinct sets of RBCs and corresponding ground truth labels were used across wavelengths.