1. Introduction

Cytology research, including cell therapy and cell biopsy, is emerging as a promising technology for early diagnosis and treatment of serious diseases, such as cancer, neurodegenerative disease, and heart disease [1–3]. In cancer treatment, for example, treatment is performed by inducing cancer cell death or making an incision to remove the cancer. After that, if cancer cells are detected through a cell viability assessment, the same process is repeated to prevent metastasis and recurrence [4]. However, there are cases of recurrence due to insufficient cancer cell death during treatment or remnants of cancer cells remaining after cancerous tissue removal [5,6]. Therefore, in order to improve the effectiveness of a cytology treatment/diagnostic method, a rapid and accurate method for examining cell viability is required. Currently, the most common cell viability assay is trypan blue exclusion assay [7]. When exposed to trypan blue dye, intact plasma membranes of live cells exclude the dye, while dead cells absorb it and appear with a blue color [8]. Trypan blue staining lacks the precision needed for many modern medical applications due to the brief 5-to-10-minute window of stability and the intensive microscopic examination required by the user [9]. These limitations result in live cells being over counted [10] and dead cells being under counted [11], both of which contribute to an over estimation of viability. Fluorescence assays improve upon the common trypan blue test by offering greater stability [12]. However, like trypan blue dye, the fluorescent dyes may affect the viability of the cells being studied. Additionally, cells cannot be reused in viability assays using labels and dyes due to permanent changes. Raman spectroscopy offers an alternative to dye-based viability assays by using lasers to measure the vibrational states of the molecules [13]. However, there remains the possibility that the high-power density of excitation lasers used in a Raman system may affect cell characteristics [14]. The above-mentioned methods additionally require biopsies, and samples for analysis are time-consuming to prepare and process.

In our previous study, we demonstrated the capability of dynamic full-field optical coherence microscopy (DFFOCM) to perform quantitative measurement of cellular dynamic activities [15]. DFFOCM is an observational method based on optical coherence tomography that enables tomographic images of optical scattering media to be observed in a real-time, label-free, non-invasive, and non-destructive manner with an axial and transverse resolution of several micrometers [15]. Building upon our previous findings, we believe that intracellular dynamic activities can be used to assess cell viability through measurements of frequency and magnitude. Meanwhile, recent machine learning (ML) technology has revolutionized the advancement of medical research. ML predictive models learn from existing data to diagnose and predict disease with high accuracy, highlighting improved decision-making rules for individual patients/samples in a clinical trial [16]. Recently, studies have been conducted to evaluate the viability of cells by combining ML and digital holographic microscopy [17–20]. In these studies, the measured phase delay was quantitatively classified by observing cells in stationary state, and cell viability was determined with 85-93% accuracy by applying ML. On the other hand, since DFFOCM can provide high temporal resolution (up to 150 Hz), it is possible to evaluate cell viability by stationary or dynamic (non-stationary) cell status [15]. We hypothesize that a cell viability assessment method combining the use of DFFOCM and ML can replace current time-consuming histological/biochemical methods.

To validate our hypothesis, first, we defined live and dead cells using the two features (frequency and magnitude) used for dynamic cell state analysis in previous studies, and we acquired data from thousands of cells for ML model training. Second, we trained our data using four well-known ML classification models. Depending on the nature of a dataset, performance of an ML algorithm can vary. In this study, we selected four commonly used ML models, including Logistic Regression [21], Random Forest [22], Support vector machine [23], and Gaussian Naïve Bayes [24], to compare the performance of these models on datasets acquired from DFFOCM. Finally, we used the trained ML models to analyze and calculate viability of cells in the process of cell death. The viability assessment by ML based on intrinsic intracellular dynamic activity is expected to be an important complement to the current histological/biochemical assessment of cell viability.

2. Methods and material

2.1 Sample preparation

HeLa cells are one of the most commonly used cells in cytology research [25] and were used in this study because of their high reproducibility. The HeLa cell line was provided by the American Type Culture Collection (Manassas, VA) and was cultured on a Snapwell membrane (3801, Corning Inc., Corning, NY) in the same manner as in previous study [15]. A total of six Snapwell membranes, including the three samples from our previous study [15], were prepared. There were approximately 30,000 HeLa cells in a Snapwell membrane. The dynamic activity of the cells was observed for 24 hours in the same field of view using a DFFOCM system.

2.2 Dynamic full-field optical coherence microscopy (DFFOCM)

The DFFOCM system used in this experiment is a commercially available Light-CT scanner (LLTech SAS, France). The schematic of the system was introduced in a previous paper [26]. This system is equipped with a 10x water immersion objective lens which has a numerical aperture of 0.3 as a standard, a broadband light source, and a high-precision axial stage, providing a field-of-view (FOV) of 1.2 mm² with a transverse resolution of 1.5 µm and an axial resolution of 1 µm. In addition, the system is also equipped with a high-speed complementary metal-oxide-semiconductor (CMOS) camera that provides a high temporal resolution of up to 6.7 ms (150 Hz). DFFOCM is a time-resolved based imaging technique by using a high-speed CMOS camera. The raw image data observed continuously in time can be efficiently selected from slow dynamic movement to fast movement of the sample according to the number of images of the sub-data that are divided to observe the dynamic activity of the sample [26]. The DFFOCM used in this study acquired 1,000 raw images of samples with a frame rate of 300 Hz and divided them into 10 sub-data (each sub-data has 100 frames) to observe the dynamic activity of cells.

2.3 Experimental procedures

A Snapwell membrane containing HeLa cells was quickly transferred from a cell’s incubator which was maintained at a temperature of 37°C and 5% concentration of CO₂ gas to a DFFOCM sample holder. The total transfer time was approximately 10 minutes. The experiment was conducted in an environment where room temperature (22°C) was maintained, without CO₂ gas supply [15]. The cells were observed for 24 hours from immediately after the sample holder was placed in the system (0-hour mark). From our previous finding [15], it was confirmed that the dynamic activity of HeLa cells observed with DFFOCM in this experiment was significantly different between 0-hour and 24-hour (details are provided in Sec. 2.4.). For model training of ML, therefore, we assumed that all the cells observed at 0-hour were alive and at 24-hour were dead. In addition, cells were observed at 15-minute, 2-hour, 4-hour, and 6-hour marks to evaluate cell death overtime using the trained model. This process was repeated six times for six cell membrane samples.

2.4 Data analysis

To observe only the dynamic activity of cells without environmental and background noise on six Snapwell membranes, cellular regions were segmented from the image data observed using DFFOCM, these regions were counted as cells, and the remaining regions were considered as backgrounds. From the acquired data, a total of 3,404 cells were observed as indicators of ML models, of which 2,164 were alive cells (0-hour) and the remaining 1,240 were dead cells (24-hour). The cell segmentation was performed by Mask Region Based Convolutional Neural Network (R-CNN) [27], implemented using libraries from TensorFlow, Keras, and PyTorch. Using the cell segmentation, the cells and background regions were separated, and frequency components for the interior and background of the cell were calculated respectively. Then, ‘mean frequency’ (${f_{\; mean}}$) was calculated after subtracting the background frequency component from the cell's frequency component. The dominant frequency was defined as the ‘${f_{\; mean}}$’, taking into account the frequency component and its power spectral density (PSD) inside the cell. To calculate the ${f_{\; mean}}$ inside the cell from the continuously observed raw data using DFFOCM, the frequency spectrum for each pixel of the cell region was calculated. The dynamic movement of the cell is a non-stationary signal, and the frequency component was observed using Welch's method [28] to analyze the small dynamic movement of the cell. Afterward, the dominant frequency was obtained from each pixel through the frequency component and PSD analyzed by Welch's method [15]. At this time, the magnitude used to train the ML was calculated by taking the average of the PSDs.

In previous study [15], we demonstrated that ${f_{\; mean}}$ and magnitude were effective parameters for evaluating cell viability states. Specifically, ${f_{\; mean}}$ and magnitude were statistically significantly different between alive HeLa cells (0-hour mark, ${f_{\; mean}}$ = 4.79 ± 0.50 Hz and magnitude = 2.44 ± 1.06) and dead HeLa cells (24-hour mark, ${f_{\; mean}}$ = 8.57 ± 0.71 Hz and magnitude = 0.53 ± 0.25) [15]. Hence, in this study, we used these two parameters to train our model. Each entry to the model included two predictor variables (${f_{\; mean}}$ and magnitude) and a binary target variable indicating cell viability. The ${f_{\; mean}}$ and magnitude values of cells from the 0-hour mark were given a ground ‘truth label’ of ‘1’ to indicate alive cells and the values of cells at the 24-hour mark were given a ground ‘truth label’ of ‘0’ to indicate dead cells. For the machine leaning, the total data was randomly partitioned into a 67:33 ratio to create a training set containing 2,280 entries and a test set containing 1,124 entries for model building and evaluation respectively. We compared model training performance using four well-known supervised ML techniques: Logistic Regression [21], Random Forest [22], Support vector machine (SVM) [23], and Gaussian Naïve Bayes [24] classifiers. Results were visualized using a confusion matrix [29] and accuracy was calculated for each model and shown in Sec. 3.1.

3. Results

Cell viability assessments were performed through the trained ML models using the two indicators (${f_{\; mean}}$ and magnitude) of cells observed from the DFFOCM. Assessment accuracy using ML were compared among the four different algorithms. Additionally, the results from the ML models were analyzed and compared to the previous results of validating the death of HeLa cells using trypan blue [15].

3.1 Classifier accuracy

To verify the accuracy of the trained models, a test dataset consisting of 714 live cells and 410 dead cells (33% of total data from each viability state) was used. Figure 1 shows the confusion matrices applied to the predictions of the four different models (Logistic Regression, Random Forest, SVM, and Gaussian Naïve Bayes) on the test dataset. A confusion matrix is a matrix allows us to visualize the performance of a supervised learning algorithm.

 

Fig. 1. Confusion matrices for the four machine learning models of the cell viability test set. The color bar indicates the number of cells.

Download Full Size | PDF

Here, the main diagonal represents ‘true’ where the ‘truth label’ of the samples and the ‘predicted label ‘ by the trained ML model match, and the remainder represents ‘false’ (error). The sensitivity and specificity of each model calculated by the confusion matrix are presented in Table 1.

Table 1. Sensitivity and specificity of confusion matrix for each model

View Table | View all tables in this article

As a result, from all the trained algorithm models, live cells were assessed with 97.89 ± 0.67% sensitivity, and dead cells were evaluated with 90.28 ± 2.34% specificity. Table 2 presents the ‘balanced accuracy’ results calculated by the specificity and sensitivity of each model. All models showed high classification performance on the test dataset ($\ge $ 93.03%), with an average balanced accuracy of 93.92 ${\pm} $ 0.86%.

Table 2. Balanced accuracy of confusion matrix for each model

View Table | View all tables in this article

3.2 Cell viability assessment

Figure 2 shows results of cell viability assessment over time by applying the Logistic Regression model to one of six Snapwell membranes data (the same dataset in our previous study [15]).

 

Fig. 2. Assessment of cell viability state over time by trained logistic regression model. After (a) 15-minute, (b) 2-hour, (c) 4-hour, and (d) 6-hour mark. Green and red labels indicate live and dead cells, respectively. The inset white scale bar represents 100 µm.

Download Full Size | PDF

Green and red labels indicate live and dead cells, respectively. The 0-hour and 24-hour mark data were not included in the assessment of cell viability over time because they were used as the ‘ground truth’ data to train the ML model. Figure 2 indicates that most cells were alive at the 15-minute mark, but then gradually died after that. At the 6-hour mark, most cells were dead. In the results of our previous study [15], HeLa cells left without CO₂ supplement at room temperature (22°C) showed dramatic changes in both ${f_{{\; }mean}}$ and magnitude between 4 and 6 hours as the cell death process progressed. Assessment by trypan blue after the 6-hour mark could infer to some extent the presence of dead cells, but it was difficult to quantify viability [15]. On the other hand, Fig. 2(d) clearly indicates that more than 83% of HeLa cells died at the room temperature without CO₂ supply after the 6-hour mark.

Table 3 numerically shows the viability of the cells shown in Fig. 2, calculated by the Logistic Regression model. Three other ML models showed similar results (data not shown in this paper).

Table 3. Viability of HeLa cells over time without CO₂ supply at 22°C

View Table | View all tables in this article

4. Discussion and conclusion

This study proposes a novel method to assess cell viability based on ML and intracellular dynamic activity. Intracellular activities, including ${f_{\; mean}}$ and magnitude of the six HeLa cell samples (total 3,404 cells) were measured for 24 hours in the same field of view using a DFFOCM system. Cell dynamic activities at 0-hour and 24-hour marks were then fed to supervised ML algorithms to train models that can distinguish live from dead cells. During the observation of the process of cell death in our experiment, we considered that cells were decomposed [30] and/or detached from the Snapwell membranes surface, which deviated from the observation region of the DFFOCM system. This resulted in a discrepancy in the counted cells between alive (0-hour; 2,164 cells) and dead (24-hour; 1,240 cells). Nevertheless, four supervised ML models (Logistic Regression, Random Forest, SVM, and Gaussian Naïve Bayes), resulted in a balanced accuracy of 93.92 ± 0.86%. The tendency in the data obtained through the DFFOCM was found to be less than 1% (standard deviation) in the performance results of the four ML models with different learning algorithms. This implies that data obtained through the DFFOCM can have high compatibility with other ML models. Though a high balanced accuracy was obtained, the cell viability classification was not perfect (100%). This could be because we assumed that all cells at the 0-hour mark are alive while all cells at the 24-hour mark are dead. After training the ML models, we used them to assign cells to a viability state and calculated the total cell viability of the sample. We observed that the cell samples had a viability of 95.59% at the 15-minute mark and gradually decreased to 52.86% at 4-hour and 16.30% at 6-hour marks. The observed gradual decrease in cell viability over time follows a natural cell death process in a non-CO₂ supplemented environment [3]. The results of the evaluation of cell death in the absence of CO₂ supplementation were clearly different between ML and trypan blue evaluations. Dead cells with blue membranes were partially observed after the 6-hour mark with trypan blue [15], but it was clearly indicated using intrinsic activity and supervised ML. The ambiguity of trypan blue discrimination may be due to the short stability of staining and non-reaction due to cellular degradation [8].

Intracellular dynamic activity learning using ML is expected to make an unrivaled contribution to clinical trials as it has the potential to predict future cell status in addition to cell death assessment. For example, it can be used for the treatment and diagnosis of various tumors in real-time, especially on complex, dynamic, and malignant cancer cells that have adaptability. Common oncology therapy consists of pathological observation of the cancer cell cycle phase at the site of tumor infection, followed by repeated physical and cytotoxic treatments. However, cancer cells remaining after treatment may develop into immune cancer cells and cause proliferation [31]. The observation of the intracellular dynamic activity using DFFOCM, based on the ML identification method is repeatable in the same field of view by analyzing the proliferation potential and dead cells in a real-time, label-free, non-invasive, non-destructive, and in-vivo manner, thereby preventing the proliferation of cancer cells at an early stage.

The current study performed a cell viability assessment on HeLa cells, which is an immortal human cell line used widely in cell research. In other cell types, the intracellular dynamic activity observed in HeLa cells may be different. To account for this possibility in future studies, other primary tumor or cancer cell lines, such as MCF-7 for breast cancer, A549 for lung cancer, and Jurkat cell line for applications in immunotherapy could be tested. Our findings can be applied to various fields requiring cellular analysis, so we anticipate that these findings will make a great contribution to medical research.