Abstract
Background/Aim: This study aimed to automate the classification of cells, particularly in identifying apoptosis, using artificial intelligence (AI) in conjunction with phase-contrast microscopy. The objective was to reduce reliance on manual observation, which is often time-consuming and subject to human error. Materials and Methods: K562 cells were used as a model system and apoptosis was induced following administration of gamma-secretase inhibitors. Fluorescence staining was applied to detect DNA fragmentation and caspase activity. Cell images were obtained using both phase-contrast and fluorescence microscopy. Two AI models, Lobe(R) and a server-based ResNet50, were trained using these images and evaluated using F-values through five-fold cross-validation. Results: Both AI models demonstrated effectively categorized individual cells into three groups: caspase-negative/no DNA fragmentation, caspase-positive/no DNA fragmentation, and caspase-positive/DNA fragmentation. Notably, the AI models’ ability to differentiate cells relied on subtle variations in phase-contrast images, potentially linked to changes in refractive indices during apoptosis progression. Both AI models exhibited high accuracy, with the server-based ResNet50 model showing improved performance through repeated training. Conclusion: This study demonstrates the potential of AI-assisted phase-contrast microscopy as a powerful tool for automating cell classification, especially in the context of apoptosis research and the discovery of anticancer substances. By reducing the need for manual labor and enhancing classification accuracy, this approach holds promise for expediting high-throughput cell screening, significantly contributing to advancements in medical diagnostics and drug development.
Blood count and cell classification are essential in blood tests and are performed automatically, to some extent (1). For fine classification, such as for apoptosis or differentiation, naked-eye observation using a microscope is inevitable. Such a classification relies on knowledge that is difficult to express or extract, i.e., on tacit knowledge, which is, thus, more difficult to transfer to others objectively in writing. Medical technology laboratories in different hospitals make every effort to unify the standard for cell classification (2). Therefore, an automated method that enables fine cell classification is desired, and some trials have been reported (3). An automated classification method also solves the problem of the observer’s exhaustion, which can lead to misclassifications. Observation for cell classification is often accompanied by cell staining using coloring or fluorescent dye. Cell classification in cytology is professionally performed by medical technologists in hospital laboratories (4). The identification of cell apoptosis is required for clinical and basic cancer research purposes (5). Accordingly, in drug discovery studies, vast numbers of experiments examining the apoptosis-triggering effects of various materials are required (6). Buffered Wright stain, Wright-Giemsa stain (a combination of Wright and Giemsa stains), and buffered Wright-Giemsa stain have been chosen thus far for microscopically classifying cells. Such staining results in cell death, making subsequent live-cell examinations impossible.
Generally, apoptosis is characterized by a series of typical morphological features, such as shrinkage of the cell, fragmentation into membrane-bound apoptotic bodies, and rapid phagocytosis by neighboring cells (7, 8). At the single cell level, the activation of a series of proteases termed caspases occurs. Caspase-3 cleaves the inhibitor of the caspase-activated deoxyribonuclease and the caspase-activated deoxyribonuclease (9). Fragmented DNA and caspase activity can be identified using fluorescence microscopy. The fluorescent stain SYBR green is, for example, used to stain DNA (10). Markers of apoptosis focusing on caspase activity are based on the interaction between the enzyme and a FITC-conjugated inhibitor. Apoptotic cell death is also determined by measuring mitochondrial membrane potential (11). These assays are complemented with fluorescent dye staining. This technology makes it possible to view dynamic processes in living cells; however, extended observation using fluorescence microscopy (both wide-field and confocal) can result in significant light energy exposure (12). Therefore, it is possible that cells experience light-induced damage that alters cell physiology, preventing intact cell observations. Neither coloring dye nor fluorescent dye methods are, therefore, suitable for observing cell morphology dynamics throughout the apoptosis process. Flow cytometry is an effective tool for observing cell characteristics (13), but unlike microscopy, cell location information is lost in this method, i.e., the geometric (position) information of cells is lost. A future high-throughput apoptosis test for drug discovery requires 1) microscopic apoptosis cell detection 2) followed by detailed analysis through cell revisit. A sample revisiting system with micrometer resolution is already realized in an astrobiology research project (14). Although the fluorescence microscope-based observation of cells is precise compared to that of phase contrast, it may stress cells both chemically (by the addition of dye) and physically (by irradiation). The observation of cells in a tactic condition is desired, and for this purpose, the acquisition of the morphological information of cells should be, for example, favored. In view of limiting the apoptotic features in morphology, the execution phase of apoptosis is characterized by cell contraction and membrane blebbing (15).
Therefore, we came up with the idea that a combination of morphology-sensitive microscopy and AI (artificial intelligence) might enable a novel, stainless way of discriminating apoptotic cells from normal ones. In other words, training AI using phase-contrast images and precise fluorescence images might result in the development of an automatic cell classification system. Phase-contrast microscopy is used for cell observation without needing a stain reagent (16, 17). Time dependent cell morphology change is observed using video time-lapse microscopy (18). A combination of scanning electron microscopy (SEM), fluorescence microscopy, and phase-contrast microscopy are used for observing cell apoptosis (19). Intact cell information is lost in SEM observation due to vacuum environment. As AI is known to classify images in ways different from the human naked eye, when AI will be able to classify the cells without dynamic (time-lapse) information, a gate for quick high-throughput cell classification could be opened. Lobe, a machine-learning model development software developed by Microsoft, operates on a PC and allows users to create their own classification system by training at least five images per label (20). A data set or training data including normal and apoptotic cell images, should be produced to teach an AI. It is essential to extract morphological features from biological cell photographs to use as features for machine learning models. In the case of adherent cells, such as Hela cells, multi-layered or adhered cell images are commonly observed, and such features make cell boundaries quite unclear. To obtain good training data with thousands of labeled cell images, suspended cells are preferred, as they appear with clear cell boundaries in the microscopic image. K562 is a suspended, human leukemic cell line used as a model of differentiation (21). We used this cell line as a model of leukemic cells because they show good dispersive characteristics when observed under microscopy. Notch activation is involved in the growth of leukemia cells. γ-Secretase inhibitors (GSIs), which block Notch activation, are reported to inhibit the growth of K562 cells via cell-cycle inhibition (22). Fluorescence microscopic images of K562 cells treated with GSI showed apparent features of apoptosis (23). The processing and classification of apoptotic cell images using a common PC application should be a good starting point before going to a sophisticated, authentic AI server.
In this study, we attempted to classify (normal (CA−/Frag−), caspase positive (CA+/Frag−), and caspase positive/DNA fragmentation positive (CA+/Frag+)) phase-contrast images of individual K562 cells using an easy-to-use AI (lobe(R)), and to examine its accuracy using fluorescence-based right answers. As both caspase activity and DNA fragmentation are essential indices of apoptosis, we focused on these two events for the classification. We also used a server AI for cell classification to compare the results produced by the two AIs.
Materials and Methods
K562 (chronic myeloid leukemia) cells were supplied by the Japanese Cancer Research Resources Bank and cultured in minimum essential medium (MEM Alpha, Life Technologies, Carlsbad, CA, USA) supplied with fetal bovine serum (FBS), at 37°C under 5% CO2. Apoptosis induction was performed by adding GSI-XXI (Compound E, purchased from Merck, Rahway, NJ, USA) 72 h after starting incubation at a final concentration of 20 μM. Cell images were acquired three days after the addition of GSI. Fluorescent staining of DNA and caspase activity was performed by the addition of SYBR Green I (SYBR™ Green I Nucleic Acid Gel Stain; Takara Bio Inc, Nojihigashi, Kusatsu, Japan; diluted 2,000-fold) and CaspACE™ (FITC-VAD-FMK, finally diluted 10,000-fold; Promega, Madison, WI, USA). Fluorescence microscopic images were captured using a 491 nm excitation/561 nm emission filter set of an inverted microscope (DMIRB, Leica Microsystems, Ernst-Leitz-Straβe, Wetzlar, Germany). Phase-contrast (using Phase plates 0, 1, and 2) images and fluorescent images of the same field were captured. Cell images were captured in JPG format using a digital still camera (α-7S SONY, Konan, Tokyo) controlled by a PC application (Imaging Edge Remote, SONY). Fluorescence images of cells with positive caspase activity were observed as relatively brighter green images as compared to the background, where there was negative activity without such brightness. The exposure time was adjusted to capture such fluorescence. Both SYBR green and CaspACE™ showed green fluorescence. This made it possible to identify DNA fragmentation and caspase activity at the same time with one single dichroic mirror-filter system. This single-color setup also helped the AI to be free from color-based classification. DNA staining was performed using different colored dyes, such as 4′,6 Diamidino 2 phenylindole (DAPI) or propidium iodide (PI) to reconfirm the morphology of DNA; finally, SYBR green was chosen. As for microscopic images, 100 fields with tens of individual cells were captured under phase-contrast (PC) and fluorescence conditions. Individual cell images were manually cropped from the field images, using PC application (“Preview” on MacOS). Finally, 1,380 pictures with a single cell were prepared and were used to train both Lobe(R) and server AI. The individual cell images were imported to Lobe(R) and labeled. The image classification model “ResNet50” on our university server, was also used for cell classification. Lobe requires manual training and testing. Therefore, Lobe was trained once using 1380 pictures. However, server-based ResNet50 is equipped with a programmable training and testing function. Therefore, we trained ResNet50 100 times by modifying picture data step by step (random rotation, trimming, rolling over, and partial erasing) in every training by a python-based program. Eighty percent of all data was used to train AIs, and the rest was used for testing. Also, all of the data was divided into five groups (20%×5 groups), and testing was performed using 5 different groups as test data (five-fold cross-validation). F-values (24) were calculated to examine the accuracy of the AI’s performance.
Results
Interestingly, we found that using fluorescence images, the cells could be classified in three categories, as shown in Figure 1. The 1380 pictures with single cells were classified manually into three classes (Figure 1) — 475 caspase negative/no DNA fragmentation (CA−/Frag−), 665 caspase positive/no DNA fragmentation (CA+/Frag−), and 240 caspase positive/DNA fragmentation (CA+/Frag+) — by examining the fluorescence images. Typical images of normal and apoptotic cells are shown (Figure 1). In Figure 1A, which represents CA−/Frag−, an undivided nucleus emitting fluorescence is observed with dark cytoplasm. As seen in Figure 1B, representing CA+/Frag−, an undivided nucleus emitting fluorescence is observed, along with a dim image of the cytoplasm. Also, in Figure 1C, several bright dots (presumably fragmented DNA) are observed instead of a large nucleus cluster, along with a dim cytoplasm image. Table I shows the F-value of the classification results obtained using the two AIs after five-fold cross-validation.
Transmitted (A, C, E) and fluorescence (B, D, F) image examples of K562 after the addition of GSI. A-B (CA−/Frag−), C-D (CA+/Frag−), and E-F (CA+/Frag+) are in the same field of view. Scale bars represent 20 μm. ISO 250, exposure time 0.4 s (for A-E), 0.2 s (for F).
Average F-score values obtained from 5-fold validation.
Discussion
Cells are directed by Notch signaling whether to proceed to a proliferative or apoptotic state (25). As GSIs are known to inhibit NOTCH signaling (26), the observed caspase activity and/or DNA fragmentation should be caused by the addition of the GSI. Chromatids can be visualized in phase-contrast images (27). Refractive index change has been reported with DNA concentration change (28). The refractive index of DNA is reported to be 1.5, while that of cytosol 1.36-1.39 (29). Chromatin condensation proceeds as apoptosis proceeds (30). Chromatin bodies, consisting mainly of DNA and histone, are reported to scatter visible light in a quantitative way (31). Therefore, AIs might be detecting the faint difference in the shade pattern in the images of cells undergoing apoptosis. Acquiring information from phase-contrast images using refractive indices has not been recommended traditionally (32), but our findings introduce the possibility of using phase-contrast images to detect apoptosis.
Why, then, did the AIs discriminate (CA−/Frag−) from (CA+/Frag−)? These two classes commonly showed no DNA fragmentation. CaspACE™ fluorophores show green fluorescence through absorbing blue excitation light. In phase-contrast images, a slight difference in the color balance inside the cell might have occurred when caspase was activated. AI might have discriminated such a difference in the phase-contrast images. Individual cells in the field images are now identified automatically by an instance segmentation model “Mask R-CNN.” All individual cells in microscopic field images are now automatically cropped so that the time required for the cumbersome work before annotation is shortened. Thus, our findings should help advance the search for anticancer materials. As mentioned, Lobe training was performed only once, and ResNet50 was programmed to repeat training. Both AIs showed similar accuracy after the first training. As AI accuracy generally approaches 1 with training, repeated training is quite effective with a limited number of cell images. Greater accuracy in the case of ResNet50 should be the result of repeated training. Even with the PC standalone AI application, support from a suitable training program should promise excellent cell classification. Such an environment is useful in the search for anticancer materials in the lab.
Acknowledgements
This study was partly supported by the funds for Bionics program, Tokyo University of Technology Graduate School.
Footnotes
Authors’ Contributions
YK and SS initiated and designed the study, and prepared this manuscript. YK and YO performed cell culture. YK performed microscopic observation, photography, and data processing. YK and HI performed AI machine learning. MK and SS provided advise on the article.
Conflicts of Interest
The Authors have no conflicts of interest to declare in relation to this study.
- Received September 28, 2023.
- Revision received October 31, 2023.
- Accepted November 6, 2023.
- Copyright © 2024 The Author(s). Published by the International Institute of Anticancer Research.
This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY-NC-ND) 4.0 international license (https://creativecommons.org/licenses/by-nc-nd/4.0).
