Boosting polyp screening with improved point-teacher weakly semi-supervised.

Area of Science:

• Medical Imaging
• Computer Vision
• Oncology

Background:

• Colon polyps are precursors to colorectal cancer, necessitating early detection.
• Current screening methods face challenges due to limited annotated data and image quality issues like blurred boundaries and low contrast.
• Existing weakly semi-supervised methods often overlook valuable intermediate features from teacher models.

Purpose of the Study:

• To develop an effective polyp segmentation method for scenarios with limited annotation data.
• To address the challenges of poor performance in polyp detection caused by data scarcity and image quality.
• To improve upon existing weakly semi-supervised learning techniques for medical image analysis.

Main Methods:

• Leveraging the combined strengths of Convolutional Neural Networks (CNN) and Transformers for robust polyp feature extraction (boundary and context).
• Implementing a novel teacher-student intermediate feature distillation method to guide student model learning.
• Utilizing point-prompt teacher models adapted for complex medical images and limited annotations.

Main Results:

• The proposed method demonstrates effective handling of scenarios with limited annotations.
• Achieved good segmentation performance for colon polyps.
• The integration of CNN and Transformer inductive biases proved beneficial for feature representation.

Conclusions:

• The developed method offers a promising solution for accurate colon polyp segmentation in resource-limited settings.
• Intermediate feature distillation enhances learning in weakly semi-supervised segmentation tasks.
• This approach contributes to improving early detection of colorectal cancer through advanced AI in medical imaging.