Childhood acute lymphoblastic leukaemia responds to standard treatment, but more targeted
drugs are needed. Patient-derived xenografts (PDX) more closely resemble patient cancers
than cell lines. PDXs do not proliferate well ex vivo without mesenchymal stromal
cells (MSC). Separation of the cell types may allow greater accuracy and insight into
patient drug responses observed in the clinic. Drugged PDX-MSC cells were stained
with a fluorescent DNA dye and imaged. After QC images were analysed by object-based
(OB) or pixel-based (PB) classification pipelines, using supervised machine learning.
Ground truth images determined the accuracy and precision of each approach. Combination
treatments were assessed using SynToxProfiler. OB classification resulted in an excellent
correlation with ground truth PDX counts, but not MSCs (R2 = 0.93, 0.36 respectively).
Overlapping pixels between ground truth and called objects gave a false positive rate
of 0.4 % for PDX and MSC, but the false negative rate was 23 % and 47 % respectively.
PB improved on cell number correlation for both cell types (0.98, 0.83), and false
positive/negative scores were reduced (PDX < 0.1 % & 15 %, MSC 0.2 % & 32 %).
