Cytopathological quantification of NORs using artificial intelligence to oral cancer screening

Abstract

Oral squamous cell carcinoma (OSCC) remains the most prevalent neoplasm of the head and neck. In recent decades, the incidence and prevalence of OSCC have not significantly changed, highlighting the critical need to develop and implement new risk assessment measures. The present study aimed to define argyrophilic proteins of the nucleolar organizer region (AgNOR) cut-off risk points by oral exfoliative cytological smears comparing specialized humans with a convolutional neural network (CNN) system AgNOR Slide-Image Examiner. This study included four experimental groups: control, exposure to carcinogens (alcohol and tobacco), oral potentially malignant disorders, and OSCC. In the first phase, 50 cells were used for AgNOR quantification. In the second phase, AgNOR quantification was established in an automated manner using an AgNOR System - Slide Examiner (captured - bounding-boxed - CNN analysis). In phase 1, the cut-off point for considering a smear as suspicious was established at 3.69 AgNORs/nucleus with sensitivity of 86%, specificity of 93%, and accuracy of 90%. In phase 2, the analysis of the intraclass correlation coefficient of AgNORs attributed to the system and human was 0.896 (95% confidence interval = 0.875-0.915; p < 0.0001), and this quantification with the CNN was 20 min compared to 67 h, considering human analysis. The AgNOR Slide-Image Examiner successfully differentiated the nuclei and accurately quantified the number of NORs in oral cytological smears. The cut-off risk point of 3.69 AgNOR/nucleus indicates a suspicious sample may contribute to improvements in oral cancer screening.

Figure 1. Scheme of cytopathological collection of oral mucosal smears and quantitative assessment of AgNOR in phase 1 by human manual quantification and in phase 2 by CNN quantification.

Figure 2. Cytosmear examples. Adequate samples for analysis: A) Cell containing 3 NORs/nucleus. B) Cell containing 5 NORs/nucleus. C) Cell containing 2 NORs/nucleus. Inadequate samples that were excluded from the study: D) Nuclear overlap and artifacts in the sample. E) Nucleus with poorly defined margin. F) Cell on the left: ill-defined cytoplasm. Cell on the right: nucleus without labeling for NORs. AgNOR. 1000×, oil immersion.

Figure 3. Quantification of NORs using the LabelMe system created by this group research (Ronnau et al.13) with segmentation using a bounding box. The bounding box technique was employed, digitally marking a rectangle with minimal dimensions to determine the nuclear area containing black dots corresponding to AgNORs. Demarcations were performed until reaching the first 50 cells well distended and not overlapping for posterior analysis by CNN. Cells bounding-boxed should present a clearly identified contour of the nuclear area. Nucleus overlapped or nuclear membrane presenting overlapping black dots were excluded. A: Bounding box around each selected nucleus. B: Identification of nuclei and NORs performed by the system.