Automatic classification of EpCAM+, Cytokeratin+ objects versus survival in castration resistant prostate cancer

Introduction: Circulating tumor cells (CTC) in patients with metastatic carcinomas are associated with poor survival and may guide therapy. CTC are morphologically heterogeneous and many research groups apply different morphological definitions. Manual assignment of CTC is therefore subjective. We evaluated automated image classification of EpCAM+, Cytokeratin 8, 18 or 19+ (CK+) objects, in stored images of castration resistant prostate cancer patients (CRPC) from the multicenter prospective IMMC38 trial (1) and compared these with manually classified CTC. Materials and methods: Included were 170 patients with CRPC and 66 healthy donors patients with CRPC. Median follow-up of the patients was 15 months. The CellSearch system was used to enrich EpCAM+ objects and stain them with DAPI, CD45 and CK. Digital images of patient samples before and after initiation of cytotoxic chemotherapy were acquired. EpCAM+, CK+ objects were classified automatically with an algorithm written in Matlab (Mathworks, Natick, MA). A graphical user interface was created in Matlab to obtain a multidimensional view of the parameters measured from the objects. Various parameters such as maximum intensity value, compactness and size were used to define gates. The data were dichotomized based on the median count for various gate combinations. The gates were optimized by maximizing the Cox Hazard Ratio for overall survival. Finally, optimized gate settings were applied on the groups of baseline patients, follow-up patients and healthy controls. Results were compared with the CellSearch CTC definition: EpCAM+ object, >4 µm, DAPI+, CK+, CD45- with a morphological appearance of a cell. Results: Automated CTC counting is fast and perfectly reproducible. Automatic count of EpCAM+, CK+ objects resulted in comparable Cox Hazard ratios for baseline and follow up samples as obtained with the CellSearch CTC definition. The most relevant parameters for gating were the presence of CK signal, their size and compactness. The absence of CD45 and the presence of DAPI were off lesser importance. Presence of signal in the FITC marker channel did not contribute to the Cox hazard ratio. Objects smaller than 4 µm (not included in the CellSearch definition) also strongly related to survival although a background was noted in healthy donors. Conclusions: Automated CTC counting has equivalent predictive value to the manual count by the CellSearch definition, but is perfectly reproducible and very fast. It allows for standardization and optimization of the CTC definitions. The automated CTC counting can now be validated on independent data sets.