Parasitemia titration results.

Logscale plots a-b,d-f contain shaded regions corresponding to nominal mean parasitemia plus or minus one standard deviation of the underlying Poisson distribution when counting N total cells, where N is varied from 100 up to 100,000 in roughly 3× increments. The nominal parasitemia in panels a-b,d-f is derived from a manual count of the titration high point using 2,214 total cells, with all remaining nominal diluted values computed by theoretical dilution factor. a) Sample parasitemia extracted from the raw 285 nm classifier results is plotted in dark blue. Post-processing using the maximum slice confidence technique is shown in blue, resulting in a reduction of the false-positive rate evidenced by a further drop in the lowest points of the curve. The light blue data show the result of human collaboration on the putative infected and lowest-confidence healthy cells. The Faster R-CNN method was used to detect cells directly on the raw images (yellow). b) The data from a) were re-plotted after compensation for recall and FPR, for an array of confidence threshold values applied to the data. c) Recall and FPR compensation values are plotted as a function of confidence threshold, which both decrease with threshold value, as healthy cells’ median confidence is on the order of 99.9%. d) Manual counting of Giemsa-stained smears by three experienced technicians are plotted with solid red curves (human 1: squares, human 2: hexagrams, human 3: pentagrams), and their arithmetic mean plotted in solid black (circles). Missing markers correspond to points where no parasites were located in the 300 cell sample. Blue data with error bars represents the data from Table 2 in [8], showing the result of extensive validation across hundreds of trained microscopists. Error bars indicate standard deviation across participants’ parasitemia estimations. The “Reference mean parasitemia” column from the cited table was used as x-values during overlay. e) The titration results were plotted for the classifier trained on commercial microscope at 405 nm. Dark green circles: raw classifier results, Green ‘+’ signs: Compensated classifier results, and light green ‘*’: after human collaboration, as described in a). f) Commercial microscope classifier results after compensation, for various confidence threshold values. Large cell count numbers permit low relative counting variance, reducing deviation from the underlying distribution. g) Compensation values for recall (top), FPR (middle), and remaining cells after thresholding (bottom) from the data in f). h) Ratiometric error of the data in panels a,b,d,e, and f) vs. nominal parasitemia. Top: Manual counting of Giemsa-stained smears (d), middle: UV Scope classifier at 285 nm (a,b), and bottom: Commercial microscope images at 405 nm (e,f). All data in h) are referenced to the legends in their corresponding raw data panels. All titration data was purely ‘test data’, acquired after classifier training and validation had been completed.