Figure 3
White blood cell counting algorithm training and validation. (a) White blood cell (WBC) counts from the image batch analysis algorithm were compared to manual counts from one rater across concentrations 0, 10, 50, 100, 200, and 300 WBCs/mm3 to determine optimal parameters for maximising specificity and sensitivity using 347 images taken by OpticLine. No cell staining was used for manual counting. Parameter optimisation was iterative, and the final selected parameters produced algorithm cell counts that had the least significant difference with manual counts, evaluated with Wilcoxon signed-rank tests (with continuity correction) and Pearson correlation coefficients. Scale bar represents 200 μm. (b) Algorithm cell centroid outputs using optimal parameters for images of WBCs in patient effluent at concentrations of 0, 100, and 300 WBCs/mm3. Raw images (top row) are masked (middle row) and cell centroids are plotted on the raw images (bottom row). Arrows highlight artifacts from the camera and moving effluent. Scale bar represents 200 μm. (c) Batches of images (n = 100) of each tested WBC concentration were run through the image batch analysis algorithm to output raw cell counts. To limit noise, raw counts were filtered by taking the interquartile range (IQR), only keeping the middle 50% of the counts. k-fold cross validation (k = 10) was performed on the filtered counts to train and test predictions for WBC concentrations. Within each iteration, 10% of the data was used for testing and the remaining 90% was used for training an ordinary least squares linear regression model. An average regression line, R2, and WBC concentration predictions per known concentrations are output. (d) Graph of the k-fold average test predictions across k = 10 iterations of the cross validation vs. actual concentrations of the effluent samples. Error bars are s.d. The predicted and actual WBC concentrations were similar, falling close to one-to-one correlation as represented by the y = x line (dotted line) with average R2 across the k = 10 iterations of 0.93.
