Multi-Task Collaborative Assisted Training Method for Grouping Fuzzy Categories Classification of Cervical Cancer...

Area of Science:

• Medical Imaging
• Computational Biology
• Oncology

Background:

• Cervical cancer poses a significant threat to women's health.
• Deep learning models have improved cervical cell classification but face limitations.
• Challenges include inter-class similarity, intra-class variability (single cells vs. clusters), and annotation accuracy.

Purpose of the Study:

• To develop a novel multi-task collaborative framework for enhanced cervical cell classification.
• To overcome key obstacles hindering current deep learning-based classification accuracy.
• To improve automated detection and diagnosis of cervical cancer.

Main Methods:

• Proposed a multi-task collaborative framework with several auxiliary branches.
• Grouping cell contrast auxiliary branch for inter-class feature learning using supervised contrastive learning.
• Multi-level cell classification auxiliary branch for 5, 3, and 2-class tasks to constrain inter-class relationships.
• Image reconstruction auxiliary branch to learn contextual features and address intra-class variations.
• Soft label distillation auxiliary branch to improve annotation consistency and accuracy.
• Auxiliary branches are active only during training, not inference.

Main Results:

• The proposed framework achieved outstanding performance on HSJCC, DSCC, and SIPaKMeD datasets.
• Effectively mitigated issues of cell category similarity and intra-class variability.
• Demonstrated superior accuracy in automated cervical cell classification compared to existing methods.
• The multi-task approach successfully addressed annotation subjectivity and accuracy concerns.

Conclusions:

• The novel multi-task collaborative framework significantly enhances automated cervical cell classification.
• The integrated auxiliary branches effectively tackle the inherent challenges in cervical cell image analysis.
• This approach offers a promising solution for more accurate and reliable cervical cancer screening.