Label-free imaging and classification of live P. falciparum enables high performance parasitemia quantification without fixation or staining
Fig 2
Classification performance of deep UV microscopy at 285 nm.
Classification was evaluated with either two, three, or four categories. For each level of granularity, confusion matrices are used to represent detailed classifier performance. Diagonal entries correspond to correctly classified instances and off-diagonal entries correspond to incorrectly classified instances. ‘Predicted labels’ represents the classifier predicted categories and ‘Human labels’ represents the human-annotated ground truth. Overall classification accuracy is shown in the bottom right cell of each matrix and is defined as the percent of all instances that were classified correctly. The precision for each category is shown along the far right column (false positive rates underneath in parentheses), while the recall is shown along the bottom row (false negative rates underneath in parentheses). Within each cell in the central colored matrix, the total number of counts is shown (bold, top), as well as the corresponding percentage of the total population (bottom). Thick red boxes indicate the rates of healthy cells classified as parasitic. The classifier was trained on a random 90% partition of the pooled dataset containing all five best focus slices, comprising 58,203 healthy, 3,344 ring, 1,566 trophozoite, and 873 schizont images. The validation was performed on the remaining 30% partition consisting of 6,467 healthy, 371 ring, 174 trophozoite, and 97 schizont images.