Automated analysis of liquid-based oral brush cytology by means of Deep Learning – Development of a screening tool for head and neck cancers based on artificial intelligence

Background Oral squamous cell carcinomas (OSCCs) clinically present with a heterogenous appearance. Early and accurate diagnosis is imperative for improving the prognosis of patients with OSCC. Exfoliative cytology is a simple, cost-sparing and non-invasive diagnostic tool for early detection of oral premalignant and malignant lesions. This study evaluated the efficacy of artificial intelligence (AI) with Deep Learning in analysing smears of suspicious oral lesions, as compared to conventional cytopathologic assessment and histopathology. The reliability of the AI was evaluated in terms of sensitivity and specificity.

Materials and Methods 57 patients with clinically suspicious lesions were selected for the study. Liquid-based brush cytology examination was performed, followed by surgical biopsy. The obtained smears were PAP-stained and cytomorphologically assessed. Ten representative 2D images of each slide constituted the basis for the AI’s Deep Learning algorithm using convolutional neural networks. The results were then compared with the histopathological diagnosis.

Results Histological diagnosis found OSCC in 53 of 57 cases and high-grade dysplasia (oral intraepithelial neoplasia 2-3) in 3 other cases. In one isolated case biopsies were inconclusive. Overall 57 out of 58 cytology samples correlated with the histopathological findings.

The AI-analysis matched these findings, with highly dysplastic and malignant cells being detected with a sensitivity of 75,95% and specificity of 94,71% respectively.

Conclusion Liquid-based brush cytology is a reliable diagnostic instrument. AI as a tool for image segmentation and classification can serve as a useful adjunct in the diagnosis and screening of oral dysplastic lesions.

Publication History

Article published online:
24 May 2022