Computed aided classification of circulating tumor cell and CK+ subclasses.

54 Background: Fluorescence microscopy imaging system (OnQView, On-Q-ity, Waltham, MA) in combination with advanced cell capture techniques (OnQChip, On-Q-ity, Waltham, MA) provides necessary sensitivity to detect circulating tumor cells (CTCs) in a blood sample. The detection process involves automatic identification of CTC candidates from the collected imagery followed by CTC subclass identification. Subclass identification process is manual and usually leads to increased sample processing time. Methods: We have developed a fully automated CTC detection and classification system allowing for substantial increase in throughput while maintaining high sensitivity and specificity. Detection is accomplished by a robust segmentation technique. A set of 25 image-based features is automatically computed for each detected candidate. Features include texture measurements, morphology measurements, multichannel intensity and contextual characteristics. All CTC subclasses as well as artifact classes are manually labeled and verified by trained imaging technologists.A hierarchy of Multi-Layer Perceptron Neural Network (MLPNN) classifiers is then trained and used to identify and reject artifacts and to identify CTC subclasses. Results: A total of 27 prostate cancer patients and 33 normal controls with two 3.75ml blood samples per patient were used to validate techniques. Probability of successful artifact rejection was achieved to be 0.78 and probabilities of subsequent successful CTC subclass identification ranged between 0.79 and 0.98 (intact CTCs = 95%; irregular CTCs = 98%; fragmented CTCs = 82%). Conclusions: A fully automated CTC detection and classification system was developed. Testing was conducted with 27 prostate cancer patients and 33 normal controls to yield an artifact rejection probability of 0.78 and CTC subclass identification probabilities of 0.79 to 0.98.