CervixFormer: A Multi-scale swin transformer-Based cervical pap-Smear WSI classification framework.

Area of Science:

• Oncology
• Medical Imaging
• Artificial Intelligence

Background:

• Cervical cancer impacts millions globally, necessitating efficient screening.
• Current computer-assisted diagnosis struggles with whole-slide image analysis and generalization.
• There is a critical need for improved cervical cancer screening algorithms.

Purpose of the Study:

• To develop CervixFormer, an end-to-end framework for assessing cervical pre-cancerous and malignant lesions.
• To enhance the accuracy and generalizability of cervical cancer diagnosis using whole-slide images.
• To overcome limitations of current algorithms in handling staining variations and subtype imaging.

Main Methods:

• Developed CervixFormer, a multi-scale Swin transformer-based adversarial ensemble learning framework.
• Utilized a self-attention generative adversarial network (SAGAN) for synthetic image generation to address class imbalance.
• Employed a multi-scale transformer ensemble for cell identification and a fusion model for final outcome generation.

Main Results:

• Achieved high recall (0.940) and precision (0.934) on a private dataset in approximately 1.2 minutes.
• Demonstrated strong generalizability across four independent public datasets.
• Showcased superior performance in classifying smear- and cell-level datasets with varying numbers of classes.
• Visualizations highlighted feature extraction from cell nucleus and cytoplasm.

Conclusions:

• CervixFormer significantly outperforms existing state-of-the-art methods in cervical cancer lesion detection.
• The framework offers improved recall, accuracy, and computational efficiency.
• CervixFormer represents a significant advancement in computer-assisted cervical cancer screening.