Tissue-specific biomarkers have been studied to identify micrometastases in bone marrow and/or peripheral blood. Many
studies, however, have shown conflicting results for sensitivity and specificity of detection, forestalling translation of
these findings into routine clinical use for prognosis or diagnosis. Genetic instability and heterogeneity of cancers may
make using an absolute set of differential expression markers difficult, if not impossible, for accurate detection of rare
cancer cells via a simple blood test. The literature is rich with examples of pathologists using morphology to identify
cancer in tissue sections. We hypothesize that morphological features based on fluorescent staining of common
subcellular compartments, in particular, the nucleus, may be useful for detection and classification. High-throughput/
high-content image cytometry and computer-automated classification can aid pathologists to find suspicious
cells, independent of biomarkers. Feature data are collected from an in vitro spiked model of breast cancer in the
circulation; prestaining with CellTracker Orange creates a gold standard for assessing cancer origin. A neural network
classifier is designed using seven nuclear morphology features thought a priori to be important for classification. With
adequate training data, sensitive and specific detection may be achieved. Neural networks may be robustly trained to
assist pathologists in detecting a wide variety of cancers.
|