Towards better laparoscopic video segmentation: A class-wise contrastive learning approach with multi-scale feature...

Area of Science:

• Computer-aided surgery
• Medical image analysis
• Machine learning

Background:

• Segmentation is crucial for computer-aided surgery systems.
• Acquiring large annotated medical datasets is challenging due to privacy concerns.
• Unsupervised learning, particularly contrastive learning, shows promise for learning from unlabeled data.

Purpose of the Study:

• To improve the accuracy of segmentation models trained on limited annotated medical data.
• To leverage classification labels to enhance feature extraction for segmentation.
• To accelerate model convergence using both segmentation and classification labels.

Main Methods:

• Utilized a multi-scale projection head for extracting image features at various scales.
• Improved positive sample pair partitioning for contrastive learning on multi-scale features.
• Trained the model simultaneously with both segmentation and classification labels.

Main Results:

• Significantly enhanced segmentation performance on the CholecSeg8k dataset, even with 1-10% labeled data.
• Achieved a superior intersection over union (IoU) score compared to existing methods.
• Demonstrated effective feature extraction for each segmentation target class.

Conclusions:

• The proposed method effectively enhances segmentation accuracy in computer-aided surgery using limited annotated data.
• Simultaneous training with classification and segmentation labels improves feature representation and convergence speed.
• This approach offers a viable solution for data-scarce medical segmentation tasks.