Fig 1.
The experiment consisted of six processes: sample preparation, microscopy, segmentation, cell cropping, labelling and training/testing. To standardise imaging conditions, we mixed LMPP and pro-B at sample preparation. On microscopy, images were obtained in four channels: DIC, BF, Ph and Alexa Fluor 594 (Alx594)-conjugated anti-CD19 antibody fluorescence. We cut out individual cells based on Ph images. Next, cells were labelled based on the intensity of the Alx594-conjugated anti-CD19 antibody fluorescence. Finally, training and testing were applied to cell images to evaluate classification performance.
Fig 2.
Workflow of cell segmentation.
(A) Original phase-contrast microscopy image. (B) Binary image. (C) Binary image after removing the regions touching the border. (D) Hole filling result. (E) Euclidean distance transform result. (F) Boundaries of cell regions superimposed on the original image.
Fig 3.
Image data preparation and cross-validation scheme.
(A) Labelling: natural logarithm of intensity histogram. The intensity was calculated by Eq (1). LMPP, pro-B and unused are colour-coded as brown, green and grey, respectively. (B) Training/testing: cross-validation with the dish as the unit. Images from the nth dish were used for testing in the nth validation. (C) Schematic of the test process in training and testing. Cell images were input into the CNN to obtain the label probability. The AUC was calculated from these label probabilities and corresponding marker labels.
Table 1.
Number of cells in Experiment 1.
Fig 4.
Experimental results from cross-validation.
In all figures, triangles, circles and squares represent validation 1, 2 and 3, respectively. In addition, blue and red marks represent data from experiments 1 and 2, respectively. (A) Cross-validation results for SVM input: cell size and CNN inputs: contour and single-channel (BF, Ph, DIC). (B) Cross-validation results for CNN inputs: combination of multiple channels, z- and time stacks of DIC. (C) AUCs are shown as a function of focal positions. The shot focus was −3.6, −1.5, 0.0, 1.2 and 2.7 μm. Zero is the reference position calibrated with the Alx594-conjugated anti-CD19 antibody fluorescence image. (D) The AUCs in DIC images are shown as a function of the cell sample numbers in CNN training while the size of the test was fixed at 10,722 cell images. (E) Variation of DIC depending on the initial random number of CNN learning.