Pattern recognition and classification of two cancer cell lines by diffraction imaging at multiple pixel distances

Highlights

• Cross-polarized diffraction images allow label-free cell classification.

• GLCM yields accurate and effective features for automated classification by SVM.

• Consistent accuracies are up to 99.8% on training and up to 99.5% on 3 test sets.

• Effects of image blur on classification have been quantitatively analyzed.

• Results indicate diffraction imaging flow cytometry as a powerful cell assay tool.

Abstract

Rapid and label-free imaging methods for accurate cell classification are highly desired for biology and clinical research. To improve consistency of classification performance, we have developed an approach of pattern analysis by gray level co-occurrence matrix (GLCM) algorithm to extract textural features at multiple pixel distances from cross-polarized diffraction image (p-DI) pairs, which were acquired with a method of polarization diffraction imaging flow cytometry using one time-delay-integration camera for significantly reduced blurring. Support vector machine (SVM) based classification was performed to discriminate HL-60 from MCF-7 cells using the GLCM features and consistency of optimized SVM classifiers was evaluated on three test data sets. It has been shown that the classification accuracy of the best performing SVM classifiers at or above 98.0% can be achieved among all four data sets for each of the three incident beam polarizations. These results suggest that the p-DI pair data provide a new platform for rapid and label-free classification of single cells with high and consistent accuracy.

Introduction

Cell classification by recognition of image patterns is of fundamental interest and can have wide applications in life science and clinics [1], [2], [3], [4], [5], [6], [7], [8]. Compared to microscopy, flow cytometry (FCM) with imaging capability allows rapid data acquisition and extraction of pattern parameters from large numbers of cells on a single-cell basis [9]. The high rate of image acquisition through FCM demands development of automated image analysis algorithms to process and analyze the big data of acquired images. Based on previous studies of coherent light scattering [10], [11], [12], [13], [14], [15], we have developed a polarization diffraction imaging flow cytometry (p-DIFC) method to measure the spatial distribution of light scattered by single cells illuminated by a linearly polarized laser beam [16], [17], [18], [19], [20], [21], [22]. Different from fluorescence imaging, the cross-polarized diffraction image (p-DI) pair data acquired by the p-DIFC method record intensity distribution of coherent light scattered by a cell due to the heterogeneous 3D distribution of intracellular refractive index and needs no cell staining. The acquired p-DI pair data are results of coherent superposition of wavefields emitted by the induced dipoles of molecules in the imaged cell and thus present speckle patterns carrying molecular and morphological information of the cell [12], [13], [15]. We have shown that the p-DI pair data can be used to accurately distinguish two cell lines of high morphological similarity derived from human T and B cancer cells and two types of prostate cells derived from cancer and normal cells [20], [22].

Conventional cell images are typically acquired with non-coherent fluorescent light through a microscope or imaging FCM which present 2D projections of a 3D object. Image segmentation is generally needed for further analysis [9]. In contrast, the speckle patterns in a p-DI pair result from superposition of coherent wavefields from all excited intracellular molecules as “digital holograms” and needs no segmentation. A pixel-based global image processing algorithm often suffices to quantify the patterns or textures of such images that can be automated for rapid processing. We have developed a gray-level-co-occurrence-matrix (GLCM) based software to quantitatively characterize textures of p-DI pair data [22], [23], [24], [25]. The GLCM parameters were determined from a matrix of elements given by the co-occurring probabilities of paired pixels separated by d as the displacement vector with |d|=1. A total of 38 parameters were obtained from each p-DI pair to form feature vectors, which were used to train support vector machine (SVM) classifiers with different kernels. With the above algorithms, we have shown that the p-DIFC method performs well for accurate and label-free cell classification through automated image pattern recognition.

Despite the attractive qualities of the p-DIFC method, however, we have found that the performances of the trained SVM classifiers are not consistent when they were applied to test sets of p-DI pairs data acquired in different runs of measurements. The causes of inconsistency may relate to the variation of diffraction patterns in different data sets, which are due to cell speed fluctuations leading to different degree of blurring for images acquired with conventional CCD cameras, positioning errors of cells relative to the focus of incident beam and the high sensitivity of extracted GLCM parameters on fine pattern changes among pairs of nearest-neighbor pixels that are unrelated to cell morphology or molecular composition. In our previous study of T versus B cell lines and PC3 versus PCS prostate cells [20], [22], SVM classifiers had to be trained and tested with the p-DI pairs acquired in the same measurement run for achieving high values of classification accuracy A. The value of A decreases markedly if the SVM classifier trained with data from one run was applied to the test data acquired in another run on a different day [22]. For example, the value of A reduces from 99.5% to 62.8% in the case of PC3 versus PCS cells. Such reduction in performance prevents application of the p-DIFC method with pre-trained SVM classifiers on p-DI data acquired later.

The current study focuses on pattern analysis of p-DI data and effect of blurring as a part of our research efforts to solve the accuracy-dropping problem. A new configuration of illumination and imaging has been developed to eliminate or significantly reduce motion blur with one time-delay-integration (TDI) CCD camera and decrease the sensitivity of image patterns of acquired data with enlarged focal spot for the incident beam [21]. With this imaging configuration we have investigated the GLCM approach of pattern analysis with d≥1 to improve the robustness of the SVM classifiers. In this report, we present the results of pattern recognition and classification on two cancer cell lines, HL-60 versus MCF-7, by the p-DI data acquired with the new imaging configuration. The dependence of p-DI parameters and classification accuracy on d has been analyzed to examine the benefits of using different values of d for improved performance. We have also blurred the measured p-DI data with no or little blurring by window smoothing to investigate the effect of blurring on classification. These results demonstrate that the new imaging configuration and GLCM analysis with different d can improve significantly the robustness of SVM based classification.