Cell-Level Free Cervical Lesion Detection in Cytology Images Via Weakly Supervised Self-Correction

Area of Science:

• Medical Imaging
• Computational Pathology
• Artificial Intelligence in Oncology

Background:

• Cervical cancer is a leading global cancer in women.
• Early detection of cervical lesions is crucial for preventing progression.
• Current deep learning methods struggle with noisy or incomplete annotations in whole slide images (WSI).

Purpose of the Study:

• To develop a novel framework for enhanced instance-level cervical lesion detection.
• To address the challenges of label noise in weakly supervised multiple instance learning (MIL).
• To improve the robustness and generalization of cervical lesion detection models.

Main Methods:

• Proposed Self-Correcting Instance Learning (SCIL), a two-stage instance-based MIL framework.
• Implemented a teacher-student architecture with a weakly supervised self-correction mechanism.
• Utilized contrastive dynamic weighting and uncertainty-based self-correction to mitigate noisy pseudo-labels.

Main Results:

• SCIL significantly improved cervical lesion detection at both patch and slide levels.
• The method demonstrated enhanced feature representation and robustness.
• Effectively overcame limitations posed by imperfect annotations in cervical cytology datasets.

Conclusions:

• SCIL offers a robust solution for cervical lesion detection using weakly supervised MIL.
• The proposed self-correction mechanism effectively handles noisy labels.
• SCIL shows promise for improving early cervical cancer diagnosis through automated image analysis.