Whole Slide Imaging-Free Supporting Tool for Cytotechnologists in Cervical Cytology

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cytopathology

Background:

• Cervical cytology is vital for detecting precancerous and cancerous lesions.
• Traditional manual microscopy is time-consuming and labor-intensive.
• Existing AI systems often require costly whole slide imaging, hindering real-time, accessible diagnosis.

Purpose of the Study:

• To develop and evaluate CYTOLONE, a novel AI model for real-time cervical cytology assistance.
• To overcome the limitations of traditional and existing AI-assisted cytology workflows.

Main Methods:

• Developed CYTOLONE using OpenAI's contrastive language-image pretraining framework, fine-tuned with hierarchical labeling.
• Integrated microscope with Apple Silicon Mac and iPhone camera for image capture.
• Evaluated model performance on classification accuracy, anomaly detection, and diagnostic categories.

Main Results:

• CYTOLONE achieved superior classification accuracy compared to existing AI models.
• High Anomaly detection accuracy (95.8%) and improved accuracy in Malignancy (92.8%), Bethesda (61.5%), and Diagnosis (57.5%) categories.
• Real-time image processing in under 0.5 seconds, with clearer diagnostic category boundaries.

Conclusions:

• CYTOLONE offers a practical, efficient AI solution for cervical cytology, suitable for resource-limited settings.
• The model seamlessly integrates with traditional workflows, supporting cytotechnologists in real-time.
• Eliminates the need for expensive whole slide scanners, improving diagnostic efficiency and accessibility.