INTRODUCTION

The oral cavity closely impacts social life, not only in chewing and swallowing but also in communication and appearance. The early detection and treatment of oral cancer, which are rare (fewer than six incidences/100,000 people) in Japan,[1] contribute to improving patients’ quality of life.

Brush cytology offers a straightforward diagnostic approach to distinguish benign from malignant conditions of the oral mucosa.[2] Cytological findings, such as alterations in the nuclear-to-cytoplasmic ratio (NCR) and nuclear staining, are commonly reported in oral cytology. However, establishing objective criteria for these cytological features has been challenging because of the subjective nature of their interpretation.

With advancements in artificial intelligence (AI), the ability to accurately distinguish between normal and abnormal cells has improved. However, the criteria used and perspectives from which these judgments have been made remain a black box. At the very least, defining the type of morphology of Class 1 cells, as diagnosed by human experts, the level of cytoplasmic brightness characteristic of Class 5 cells, and the typical ranges for brightness and intensity values is essential for the integration of AI into cytological diagnostics.

We therefore aimed to define parameters of normal cytological morphology. This study focused on the following eight criteria for cellular atypia: Anisokaryosis, nuclear pleomorphism, anisocytosis, and cellular pleomorphism. Increased NCR, atypical mitotic figures, prominent nucleoli, and hyperchromatic nuclei were also observed.[3]

Focusing on the nucleus, we observed a homogeneous appearance in normal cells, whereas malignant cells exhibited heterogeneous intensity. In this study, we also attempted to evaluate both the NCR and nuclear heterogeneity.

MATERIAL AND METHODS

Cytological specimens from seven patients (three males and four females; average age, 62.4 years old) diagnosed with tongue mucosal diseases at our hospital between 2020 and 2023 were examined. Sixty cells from each Papanicolaou (Pap) Class 1, 3, and 5 were analyzed.

Laboratory procedures and data preparation

Oral exfoliative cytology was performed.[3,4] Cell samples were collected using the Rovers® Orcellex® Brush (made by Rovers Medical Devices B.V., Oss, The Netherlands), and the brush was rotated (10 times) during sample collection. Scraped brush was rapidly smeared onto glass slides. After fixed in 95% ethanol, Pap stain was performed.[5,6]

A Nikon ECLIPSE Si (made by Nikon Instruments Inc., Tokyo, Japan) with a ×40 lens (Nikon CFI E Plan achromatic, 0.65/0.65 mm; Tokyo, Japan) and an attached camera (Nikon Digital Sight 1000; Nikon Instruments Inc., Tokyo, Japan) were used to take photographs. The camera image sensor was a 1/2.8-inch color complementary metal-oxide semiconductor (CMOS). The CMOS measures 5.57 × 3.13 in area. The number of recorded pixels was 1920 × 1080, and the International Organization for Standardization (ISO) sensitivity was equivalent to that of ISO 150.

The microscope was installed in an environment without windows but with a fixed light source. Each imaging experiment was performed under identical conditions, using preset settings. Squamous epithelial cells that curled and overlapped with other cells or debris were excluded, because these factors can affect the brightness values (BV). Images were acquired at ×40 magnification and stored in tagged image file format. The area and brightness of the nuclei and cells within these images were quantified using ImageJ (National Institutes of Health, USA).[7]Figure 1 shows the selection of the cell nuclear regions. BV ranging from 0 to 255 represents the gray scale, where zero corresponds to black and 255 to white. The BVs were calculated with red-green-blue (RGB) images. The RGBs were converted to BV using the formula BV = (R + G + B)/3.[8] The absence of units and the variability in BV reflect the heterogeneity observed within individual nuclei and cells.

Figure 1: Example of nuclear region outlining (Scale bar: 10 μm).

Export to PPT

Patient consent

This study adhered to the ethical guidelines (established by the Ministry of Education, Culture, Sports, Science, and Technology of JAPAN).[9] No active invasion or intervention by human participants was conducted for research purposes. Since certain cases utilized historical data, written informed consent was secured from the patients. As certain individuals may have faced difficulties in providing informed consent, only those capable of doing so provided written informed consent. In addition, participants who could not be contacted were offered the opportunity to opt out of the study. Specifically, a summary of the research was published on the institution’s website, thereby granting participants the right to freely decline participation.

Statistical analysis

Easy R (EZR) in R Commander version 1.63 is used for statistical analysis.[10] The Steel–Dwass test (SDT) with the significance level set at 0.05 is applied for multiple comparisons.

RESULTS NCRs

The distribution of the NCRs for each Class is presented in the box plots in Figure 2. The mean, five-number summary, and standard deviation of the NCRs for each Class are summarized in Table 1. Significant differences were observed in the NCRs among the Classes (P < 0.05). Reference values for normal cells were established: Mean NCR of 0.031 ± 0.015 for Class 1, 0.074 ± 0.048 for Class 3, 0.307 ± 0.170 for Class 5. In Class 1, half of the data for the N/C ratio fell between 0.02 and 0.036, while the BV of the nucleus ranged from 90.17 to 116.67, and those of the cytoplasm ranged from 126.68 to 151.51.

Figure 2: Nuclear-to-cytoplasmic ratio (NCR) in each Class. The mean and five-number summary is shown in Table 1. As the Class number increases, the range of the NCRs widened. ✶P < 0.05.

Export to PPT

Table 1: Mean and five-number summary: Nuclear-to-cytoplasmic ratio, brightness values of the nucleus and the cytoplasm.

Class NCR BV of nucleus BV of cytoplasm 1 3 5 1 3 5 1 3 5 Max 0.097 0.247 0.780 145.16 163.40 183.54 172.66 179.77 155.07 75% 0.036 0.098 0.420 116.67 121.77 82.73 151.51 153.99 110.91 Mean 0.031 0.074 0.307 105.17 108.73 71.36 139.82 139.38 96.33 Median 0.029 0.060 0.315 104.90 106.39 69.17 141.27 141.23 97.09 25% 0.020 0.038 0.171 90.17 97.86 55.64 126.68 124.66 80.01 Min 0.009 0.017 0.016 68.90 49.91 28.11 99.10 91.25 31.65 Standard deviation 0.015 0.048 0.170 BV

The distributions of nuclear and cytoplasmic intensity values are illustrated by box plots in Figures 3 and 4, respectively. The mean and five-number summaries for each Class are presented in Table 1.

Figure 3: Brightness values of the nucleus. The mean (x) and 5-number summary are shown in Table 1. As the Class number increased, the range of BV widened. ✶P < 0.05.

Export to PPT

Figure 4: Brightness values (BV) of the cytoplasm. As the Class number increased, the range of BV widened. ✶P < 0.05.

Export to PPT

For nuclear BV, about half of the data were found between 90.17 and 116.67 for Class 1, between 97.86 and 121.77 for Class 3, and between 55.64 and 82.73 for Class 5. Similarly, approximately half of the data for cytoplasmic BV were found between 126.68 and 151.51 for Class 1, between 124.66 and 153.99 for Class 3, and between 80.01 and 110.91 for Class 5. Regarding the range of maximum and minimum values, i.e., the range of possible values increased with Class number, the spread between maximum and minimum nuclear BV values expanded across Classes: 68.90–145.16 for Class 1, 49.91–163.40 for Class 3, and 28.11–183.54 for Class 5.

Pairwise comparisons revealed significant differences in the intensity values between Classes for both nuclear and cytoplasmic measurements (P < 0.05). This result is consistent with the finding that there is little variation in the internal properties of the nuclei in Class 1; however, as atypicality increased, the nuclei became more concentrated, and darkly stained areas appeared, thus resulting in variation.

DISCUSSION

As a simple yet effective technique, cytology facilitates early detection of oral cancer and assessment of potentially malignant oral disorders. Furthermore, the accessibility of oral cytology in U.S. dental clinics has facilitated early cancer detection and improved patient outcomes.[2] From the Vital Statistics of Japan, National data indicate an upward trend in deaths from malignant neoplasms of the lip, oral cavity, and pharynx, rising from 5,066 in 2000 to 8,586 in 2023.[11]

The adoption of oral cytology in Japan has been slower than the incorporation in other countries. Data from the National Cancer Center, Japan, revealed 23,671 new patients with oral and pharyngeal cancer in 2019 (16,463 men and 7,208 women), with 7,827 deaths (5,547 men and 2,280 women) in 2020. Furthermore, the 5-year relative survival rate from 2009 to 2011 was 63.5% (60.7% for men and 69.4% for women, including those with pharyngeal cancer), and the 5-year overall survival rates for tongue cancer were 93%, 77%, 61%, and 50% for stages I, II, III, and IV, respectively.[12]

Despite being easily examined visually and by palpation, the oral cavity is associated with persistently high oral cancer mortality due to delays in diagnosis. This is due to the lack of pathologists and reliance on the experience of diagnosticians. To address this problem, we believe that an AI-based diagnostic support system will contribute to reducing the workload and minimizing human error during the diagnosis of oral cancers.

The application of AI in pathology has advanced rapidly.[13] Although histopathology has been a primary focus,[14-16] AI-driven tools have also been adopted for cytology.[17] In particular, cervical cytology has benefited from the development of AI-based systems that employ supervised learning to detect abnormalities using large annotated image datasets.[18] We previously investigated the application of AI in oral cytology.[19,20] However, the creation of a sufficiently large and representative dataset for rare cancers, such as oral cancer (with an incidence rate below six/100,000 people), presents a significant challenge. Notably, the limited sample size has increased the risk of dataset bias.[21] Furthermore, our findings highlight the presence of background bias in the dataset.[22,23] The influence of these biases on performance metrics, such as the accuracy, of AI models cannot be overlooked. Consequently, the diagnostic performance of anomaly detection algorithms is limited for small datasets.

Previous investigations of oral epithelial cell morphology have yielded divergent results. While some studies have calculated the NCR using cellular and nuclear diameters,[24-27] our study employed a more precise approach that involved the measurement of cellular and nuclear areas.[28-30] Although diameter-based measurements of nuclei and cells are simple, they are susceptible to bias due to variations in the region of interest. Given the morphological diversity of cells, which often deviates from a perfectly circular shape, we posit that area-based comparisons provide a more robust metric.

Table 2 shows previous reports about NCR. While the classification of patients has varied across studies, normal cells have been reported to have NCRs within the range of 0.019–0.031.[28-33] In our study, half of the data for normal cells (Class 1) fell between 0.020 and 0.036, with a mean of 0.031 ± 0.015, which suggests that the reference range lies around these values. Similarly, for Class 3, which warrants biopsy, half of the data for the NCR fell between 0.038 and 0.098, with a mean of approximately 0.074 ± 0.048. A high NCR warrants prompt investigation and treatment planning for malignancies. Furthermore, focusing on the internal properties of the nucleus using BV, we found that 50% of the data for Class 1 fell within the range of 90.17–116.67, Class 3 within 97.86–121.77, and Class 5 within 55.64–82.73. Finally, the SDT revealed significant differences among the Classes, which suggests that classification based on nuclear BV is feasible [Table 3].

Table 2: Nuclear-to-cytoplasmic ratios reported in the literature.

Authors Classification Median Range/SD Aktunc, et al.[28] – 0.03 0.004–0.12 Caruntu, et al.[29] Superficial cell – 0.01–0.10 Intermediate cell – 0.05–0.43 Tumoral cell – 0.36–2.34 Nivia, et al.[30] Group 1 (non-tobacco users) 0.019 ±0.019 Dagli, et al.[31] – 0.02353 ±0.0035 Hashemipour, et al.[32] – – 0.021–0.028 Patel, et al.[33] Males 0.031 0.006 Females 0.031 0.005

Table 3: Results of the Steel–Dwass test. We identified statistically significant differences in all parameters except for nuclear brightness between Class 1 and Class 3.

Nuclear-to-cytoplasmic ratio BV of the nucleus BV of the cytoplasm t P-value t P-value t P-value Class 1: Class 3 6.597249 1.26E-10 0.9755806 0.592 7.559301 1.40E-13 Class 1: Class 5 10.57616 2.89E-14 8.3765543 3.10E-14 9.805033 3.09E-14 Class 3: Class 5 10.482788 2.68E-14 9.3685082 2.89E-14 3.796812 4.31E-04

According to the 2017 World Health Organization classification, oral lesions are divided into five categories:

Negative for intraepithelial lesions or malignancy; oral low-grade squamous intraepithelial lesions (OLSIL) or low-grade dysplasia; oral high-grade squamous intraepithelial lesions (OHSIL) or high-grade dysplasia; squamous cell carcinoma; and lesions of uncertain neoplastic potential.[33] Patients with negative findings are followed up in dental clinics, whereas those with OLSIL require biopsy confirmation and subsequent follow-up by oral and maxillofacial surgeons. OHSIL and squamous cell carcinoma require treatments similar to those used for cancer, such as excisional biopsies. Furthermore, nuclear features are particularly important in differentiating OLSIL from OHSIL.[34] This study was conducted to establish reference values for normal cells, while future work should focus on the further quantification of OLSIL and OHSIL.

SUMMARY

This study established quantitative reference values for normal oral epithelial cells based on NCRs and BV, and our findings may serve as useful indicators of cellular atypia. Furthermore, differences in the NCR and BV between normal and abnormal cells may be useful as indicators for the early detection of oral cancer.

The findings provide a basis for integrating AI methodologies into oral cytology practice. However, more research is necessary to update the diagnostic criteria and create more robust AI models. By addressing the limitations of the small dataset size and background bias, future studies will contribute to improving the accuracy and reliability of AI-assisted oral cytology.

AVAILABILITY OF DATA AND MATERIALS

The participants of this study did not give written consent for their data to be shared publicly, so due to the sensitive nature of the research, supporting data are not available.

ABBREVIATIONS

AI: artificial intelligence

BV: brightness values

EZR: Easy R

NCR: nuclear-to-cytoplasmic ratio

OHSIL: oral high-grade squamous intraepithelial lesions

OLSIL: oral low-grade squamous intraepithelial lesions

Pap: Papanicolaou

RGB: red-green-blue

SDT: Steel–Dwass test

AUTHOR CONTRIBUTIONS

EM: Conceptualized the study, wrote the initial draft of the manuscript; TD, HN, TM, SI, AN, and AT: Data curation and analysis; MH, SW, TO, SY, and HN: Statistical analyses. All authors directly accessed and verified all of the data. All authors subsequently read and revised the initial draft and read and approved the final version. All authors had access to all the data in the study and have final responsibility for the decision to submit for publication. All authors are eligible for ICMJE authorship.