Cervical Cell Image Classification-Based Knowledge Distillation.

Area of Science:

• Medical Imaging
• Computer Science
• Artificial Intelligence

Background:

• Deep learning models for cervical cell classification face challenges with parameter redundancy and poor generalization.
• Accurate classification of cervical cytology smear images is crucial for early disease detection.

Purpose of the Study:

• To develop an improved deep learning method for intelligent cervical cell classification.
• To address parameter redundancy and enhance model generalization in cervical cytology image analysis.

Main Methods:

• A novel method combining transfer learning and knowledge distillation was established.
• A multi-exit classification network with embedded global context modules was used as the student network.
• A self-distillation approach fused contextual information, guiding shallow classifiers with deep ones.

Main Results:

• The method achieved high performance on the SIPaKMeD dataset with 98.52% accuracy, 98.53% sensitivity, 98.68% specificity, and 98.59% F-measure.
• The model demonstrated strong generalization capabilities, verified on a natural image dataset.
• The self-distillation strategy effectively fused contextual information for improved classification.

Conclusions:

• The developed transfer learning and knowledge distillation method significantly enhances cervical cell classification.
• The multi-exit network with self-distillation offers a robust solution for intelligent cytology image analysis.
• The proposed approach shows promise for improving diagnostic accuracy in cervical cancer screening.