Computer-assisted quantitative framework for whole slide cervical image grading driven by time series features.

Area of Science:

• Computational pathology
• Artificial intelligence in oncology
• Medical image analysis

Background:

• Current ThinPrep cytologic tests for cervical cancer diagnosis rely on manual screening, which can be subjective and lack robustness.
• There is a need for more objective and accurate methods to aid pathologists in grading cervical lesions.

Purpose of the Study:

• To introduce a two-stage quantitative detection framework for whole slide cervical images to assist pathologists in lesion grading.
• To enhance the precision and robustness of cervical cell classification and detection.

Main Methods:

• Utilized a You Only Look Once (YOLO) network with an attention module and multi-scale feature fusion for cell classification.
• Incorporated quantitative DNA description and the Matthew effect for refined diagnostic contributions and cell proliferation assessment.
• Extracted time series features and global smear information for enhanced detection robustness and resistance to false classifications.

Main Results:

• Achieved high cervical cell classification precision (0.8647) and true positive rate (95.8%).
• Demonstrated smear-level accuracy, sensitivity, and specificity of 0.9193, 0.9285, and 0.9234, respectively.
• Attained grading accuracy comparable to professional pathologists for cervical intraepithelial neoplasia assessment.

Conclusions:

• The proposed AI framework significantly improves cervical cancer detection and grading accuracy.
• Time series features are crucial for cervical cancer detection, correlating with patient physiological states.
• The model can be seamlessly integrated into existing diagnostic systems, enhancing screening efficiency and robustness.