A Multi-Task Feature Fusion Model for Cervical Cell Classification.

Area of Science:

• Medical Imaging and Diagnostics
• Computational Biology and Bioinformatics
• Artificial Intelligence in Healthcare

Background:

• Cervical cell classification is vital for automated cervical cancer screening.
• Current deep learning models show promise but require further performance enhancement for practical clinical application.

Purpose of the Study:

• To develop an advanced multi-task feature fusion model for improved cervical cell classification.
• To enhance the accuracy and reliability of automated cervical cancer screening systems.

Main Methods:

• Proposed a multi-task learning model incorporating an auxiliary task (manual feature fitting) and two main classification tasks (2-class and 5-class).
• Implemented low-layer feature fusion to integrate auxiliary task information into the main classification tasks.
• Introduced a novel label smoothing technique based on cell category similarity to incorporate inter-class information and mitigate label noise.

Main Results:

• The proposed model demonstrated superior performance compared to state-of-the-art methods on the HUSTC and SIPaKMeD datasets.
• Achieved high sensitivity (99.82%) and specificity (98.12%) for the 2-class classification task on the HUSTC dataset.
• The multi-task approach and label smoothing effectively enhanced classification accuracy and robustness.

Conclusions:

• The developed multi-task feature fusion model significantly improves cervical cell classification accuracy.
• The method shows strong potential for reducing cytologist workload in cervical cancer screening.
• The integration of auxiliary tasks and advanced label smoothing offers a promising direction for medical image analysis.