Abstract 4175: Automated classification of circulating tumor cells optimized using clinical outcome of castration resistant prostate cancer patients

Abstract

Abstract Introduction: Circulating tumor cells (CTC) in patients with metastatic carcinomas are associated with poor survival and can be used to guide therapy. Classification of CTC can however be subjective as they are morphologically heterogeneous. Method: Digital images from 175 castration resistant prostate cancer (CRPC) patients and 68 healthy donors acquired using the CellSearch(TM) system during the multicenter prospective study IMMC38 (de Bono JS, et al. ClinCanRes 2008;14(19):6302-09) were used to define CTC with algorithms developed in Matlab 2009a. Objects were segmented in the images of Cytokeratin-PE, DAPI, CD45-APC, a control channel, PE+DAPI, and the maximum intensity profile of the channels. The patients were dichotomized on the median number of objects. The Cytokeratin-PE channel returned the highest Cox Hazard Ratio (HR) and was chosen for feature extraction. Features were measured from objects in each channel. For each feature, the patients were again dichotomized on the median number of objects included after thresholding. Using only samples before initiation of treatment (baseline), the standard deviation of Cytokeratin-PE signal, the size of the objects, the peak value of the DAPI signal and the peak value of CD45-APC signal had the highest impact on the HR and a low correlation with each other. They were chosen as the features for CTC classification. Classifiers representing different morphological CTC definitions were created from these features and tested using different value ranges. Bootstrap aggregating was used to determine the robustness of the classifier. To arrive at the optimal classifier the first samples after initiation of therapy from patients (follow-up) and control samples were included in the analysis. CTC definitions to be selected were to have the most statistically significant association with overall survival (HR), and a low total count in the control samples. Results: Programmed automated computerized CTC image analyses generated CTC definitions that resulted in a HR of 3.1 for baseline and 5.2 for follow-up samples. These results were equivalent to the laborious and time-consuming manual identification of CellSearch CTC by trained operators which had resulted in HR of 3.2 and 4.6 respectively. Overall, median CTC count was 7 for baseline and 3 for follow-up samples. A total of 9 CTC were found in 68 control samples. Critically, processing of a CTC blood sample was very rapid, taking approximately 1 minute per patient. Conclusions: We have automated CTC counting utilizing a classification that optimally dichotomized castration resistant prostate cancer patient based on clinical outcome. This method is reproducible and rapid, allowing standardization of CTC counting. Citation Format: {Authors}. {Abstract title} [abstract]. In: Proceedings of the 102nd Annual Meeting of the American Association for Cancer Research; 2011 Apr 2-6; Orlando, FL. Philadelphia (PA): AACR; Cancer Res 2011;71(8 Suppl):Abstract nr 4175. doi:10.1158/1538-7445.AM2011-4175