Improving cervical cancer classification with imbalanced datasets combining taming transformers with T2T-ViT

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Cervical cancer screening relies on accurate cell classification, but faces challenges with limited, imbalanced, and low-quality public datasets.
• Existing Convolutional Neural Network (CNN) models often overfit due to these data limitations.

Purpose of the Study:

• To develop a cervical cell image generation model (CCG-taming transformers) to create high-quality, balanced datasets.
• To improve cervical cancer cell classification accuracy using advanced transformer architectures.

Main Methods:

• Proposed a CCG-taming transformers model with enhanced encoder structures (SE-block, MultiRes-block) and Layer Normalization.
• Implemented SMOTE-Tomek Links for data balancing and Tokens-to-Token Vision Transformers (T2T-ViT) with transfer learning for classification.
• Applied the model to three public datasets: liquid-based cytology Pap smear (4-class), SIPAKMeD (5-class), and Herlev (7-class).

Main Results:

• Achieved high classification accuracies: 98.79% (4-class), 99.58% (5-class), and 99.88% (7-class).
• Generated synthetic images closely resembling source data, with averaged IS of 3.75, FID of 0.71, Recall of 0.32, and Precision of 0.65.
• Demonstrated the model's effectiveness in improving classification and generating high-quality, balanced cervical cell image datasets.

Conclusions:

• The CCG-taming transformers model effectively addresses data scarcity and imbalance issues in cervical cell classification.
• This work represents the first application of transformers for both generation and recognition of cervical cell images.
• The developed model aids gynecologists in diagnosing cervical cancer stages and improves cervical cancer research through enhanced datasets.