UniCAS: A foundation model for cervical cytology screening.

Area of Science:

• Computational pathology
• Digital cytology
• Artificial intelligence in healthcare

Background:

• Cervical abnormality screening is crucial but hindered by large whole slide image (WSI) sizes, making manual examination laborious.
• Existing deep learning models for cervical cytology face challenges with morphological diversity and require task-specific designs, fragmenting clinical workflows.

Purpose of the Study:

• To introduce UniCAS, a versatile cytology foundation model designed for efficient multi-scale analysis of cervical abnormalities.
• To demonstrate UniCAS's capability in handling diverse pathological conditions and patient demographics within cervical cytology.

Main Methods:

• Pre-training UniCAS on a large dataset of 48,532 cervical WSIs.
• Integrating a multi-task aggregator for slide-level diagnosis within the UniCAS framework.
• Evaluating UniCAS performance across slide-level diagnosis, region-level analysis, and pixel-level image enhancement.

Main Results:

• UniCAS achieved state-of-the-art performance in various clinical analysis tasks.
• Achieved high area under the curve (AUC) values: 92.60% for cancer screening, 92.58% for candidiasis testing, and 98.39% for clue cell diagnosis.
• Reduced diagnostic time by 70% compared to conventional methods, demonstrating significant efficiency gains.

Conclusions:

• UniCAS establishes a new paradigm for automated cervical cytology, enabling efficient multi-scale analysis.
• The model bridges the gap between computational pathology and clinical diagnostic workflows, enhancing efficiency and accuracy.
• This foundation model offers a unified approach to diverse diagnostic tasks in cervical cytology.