Evaluation of a Methodology for Automated Cell Counting for Streak ModeImaging Flow Cytometry

Abstract

Identification of Circulating Tumor Cells (CTCs) has shown promising clinical applications, but since CTCs are found in very low concentration in blood large sample volumes are needed for meaningful enumeration. This issue impedes the analysis of CTCs using standard flow cytometry due to its low throughput. To address this issue, a high throughput microfluidic cytometer was recently developed using a wide field flow- flow cell instead of the conventional narrow hydrodynamic focusing cells (used in traditional flow cytometry) enabling analysis of large volumes at lower flow rate. This wide-field flow cytometer adopts a technique known as “streak photography” where exposure times and flow velocities are set such that the particles are imaged as short “streaks”. Since streaks are imaged with large number of pixels, they are easily distinguished from the noise which appears as “speckles” increasing the detection capabilities of the device, making it more suitable for analysis using current low sensitivity, high noise webcams or mobile phone cameras. The non-stationary nature of the high noisy background found in streak cytometry introduces additional challenges for automated cell counting methods using traditional cell detection techniques such TLC, CellProfiler, CellTracker and other tools based in traditional edge detection (e.g., Canny based filters) or manual thresholding. In order to address this issue, we developed a new automated enumeration approach that does not rely on edge detection or manual thresholding of individual cells, rather is based in image quantizing, morphological operations, 2D order-statistic filtering and decisions rules that take into account knowledge of the structure and expected location of the streaks in consecutive frames. We evaluated our approach comparing it with two current methods representing the major computational modalities for cell detection: CellTrack (based in edge detection) and MTrack2 (based in manual thresholding). Samples of 1 cell/mL nominal concentration were analyzed in batch size of 30 mL at flow rate of 10 mL/min and imaged at 4 frames per second (fps), the files were saved in uncompressed AVI format files. The cells were annotated and the signal to noise ratio (SNR) was calculated. For samples with average SNR greater than 4.4 dB, our method achieved a sensitivity of 91% compared to CellTrack (60%) and MTrack2 (71%). The True Positive Rate (TPR) of cells detected was 0.93 for our method compared with 0.80 for Mtrack2 and 0.29 for CellTrack. This demonstrated the ability of the algorithm to count rare cells with high accuracy for concentrations of 1 cell/mL with SNR greater than 4.4 dB. This cell counting capability will enable to automate low cost imaging flow cytometry based on CCD detector and the expansion of cell-based clinical diagnostics in resource-poor settings.