Classification of cancer cells using computational analysis of dynamic morphology

Highlights

• It is shown that cancer cells, and especially metastatic tumor cells, show very distinctive morphological behavior compared to their healthy counterparts on aptamer functionalized surfaces.

• The approach presented analyzes the optical data and quantifies the cell morphology for an instant detection of cancer cells.

• This can contribute to early cancer diagnosis.

• Three classifier models, Support Vector Machine (SVM), Random Forest Tree (RFT), and Naïve Bayes Classifier (NBC) were trained with known dataset.

• All classifier models detected cancer cells with an average accuracy of least 82%.

Abstract

Background and Objective

Detection of metastatic tumor cells is important for early diagnosis and staging of cancer. However, such cells are exceedingly difficult to detect from blood or biopsy samples at the disease onset. It is reported that cancer cells, and especially metastatic tumor cells, show very distinctive morphological behavior compared to their healthy counterparts on aptamer functionalized substrates. The ability to quickly analyze the data and quantify the cell morphology for an instant real-time feedback can certainly contribute to early cancer diagnosis. A supervised machine learning approach is presented for identification and classification of cancer cell gestures for early diagnosis.

Methods

We quantified the morphologically distinct behavior of metastatic cells and their healthy counterparts captured on aptamer-functionalized glass substrates from time-lapse optical micrographs. As a proof of concept, the morphologies of human glioblastoma (hGBM) and astrocyte cells were used. The cells were captured and imaged with an optical microscope. Multiple feature vectors were extracted to quantify and differentiate the complex physical gestures of cancerous and non-cancerous cells. Three different classifier models, Support Vector Machine (SVM), Random Forest Tree (RFT), and Naïve Bayes Classifier (NBC) were trained with the known dataset using machine learning algorithms. The performances of the classifiers were compared for accuracy, precision, and recall measurements using five-fold cross-validation technique.

Results

All the classifier models detected the cancer cells with an average accuracy of at least 82%. The NBC performed the best among the three classifiers in terms of Precision (0.91), Recall (0.9), and F₁-score (0.89) for the existing dataset.

Conclusions

This paper presents a standalone system built on machine learning techniques for cancer screening based on cell gestures. The system offers rapid, efficient, and novel identification of hGBM brain tumor cells and can be extended to define single cell analysis metrics for many other types of tumor cells.

Introduction

Over the past few years, numerous efforts have been made to detect tumor cells for early cancer diagnosis. Early detection and treatment of cancer has very high survival rates and dramatically improved quality of life. The detection of cancer cells is challenging due to their rare presence in blood samples at early stages. In a similar vein, the identification of metastatic and aggressive cells is very important for cancer staging. Several methods have been employed to capture and identify tumor cells based on their physical, mechanical, and chemical properties (e.g., size, deformability, electrical polarizability) [1]. For example, tumor cells have been recognized based on their biophysical properties using microfiltration, atomic force microscopy, micropipette aspiration, and micropore devices [2], [3], [4], [5], [6], [7]. The distinct electrical charge of tumor cells have also been utilized for electrophoresis-based separation [8]. In the past decade, ligand-based tumor cell capture platforms have been widely used for highly sensitive and selective applications [9], [10], [11], [12], [13]. Such affinity-based methods have been explored to selectively tag magnetic and fluorescent particles for cancer screening [14], [15], [16]. Although these techniques have their own advantages, these methods often require complex and time-consuming post-capture analysis for further verification. There is huge disparity in knowledge for rapid, low-cost, and highly sensitive platforms for cancer screening.

Various cell screening methods have been demonstrated to analyze static images of cells, and a significant amount of work has been reported on automated high-throughput microscopy with rapid image acquisition [17], [18], [19]. Several fluorescently tagged biopsy samples have been imaged and analyzed to classify multiple cell phenotypes and morphologies [20], [21], [22]. Nevertheless, static image-based classification techniques are limited to analyze cell composition only and cannot capture the behavior of cells. Therefore, time-resolved live-cell imaging is the way to go to analyze complex and dynamic cellular processes. It has already been shown that the morphology of cancer cells correlates to their gene expression profile [23]. Fluorescence time-lapse imaging has been used to investigate complex dynamic processes such as cell division, cell motility, and intracellular trafficking [24], [25], [26]. We have reported in our previous works that tumor cells show distinctive morphologies on bio-functionalized substrates [27], [28], [29], [30]. Visual inspection of morphological dynamics from large-scale data is very time-consuming, labor-intensive, and unreliable. Even a well-trained pathologist would have to spend a great deal of time to process and analyze the data from simple biological assay. The overwhelming size of data motivates the design of machine learning approaches for the classification of cell phenotypes, genetic sequences, protein expressions, and their physical properties [31], [32], [33], [34], [35]. The ability to process and learn from a large number of time-resolved images and a comprehensive analysis of cell gestures are reported here. It is a unique contribution towards effective detection and classification of cancer cells.

This paper presents a predictive computational framework to diagnose cancer by interpreting complex dynamic behavior of cells from microscopic time-lapse images. We have defined and extracted 50 characteristics features to quantify the intricate gestures of cells. A supervised machine learning algorithm was implemented to create a classification model that identified metastatic tumor cells derived from solid-biopsy of a cancer patient. The healthy counterparts of tumor cells, astrocytes, were also incubated on an aptamer-functionalized glass surface. We incorporated the temporal context into the annotation scheme to define the gestures of cells. Cell gestures were quantified in terms of features such as cell geometry, cellular protrusions, and morphological changes in cellular boundary over time. The feature vectors were combined in a dataset to train and validate a predictive model. The approach was demonstrated to selectively detect metastatic human glioblastoma (hGBM) cells and astrocytes with an average accuracy of 85%. The approach is equally valid for large-scale applications towards the detection of tumor cells without labeling the cells. This is a novel system to detect cancer cells from cell gestures and can be readily used by pathologists and life science researchers to study cell behavior for disease diagnosis.