Reducing annotating load: Active learning with synthetic images in surgical instrument segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Surgical Technology

Background:

• Accurate instrument segmentation in endoscopic minimally invasive surgery is crucial but challenging due to complex visual data.
• Deep learning models excel at this task but typically require large annotated datasets, creating a significant labeling workload.

Purpose of the Study:

• To develop an efficient framework for training deep learning models for surgical instrument segmentation.
• To reduce the manual annotation effort required for creating high-performance segmentation models.

Main Methods:

• An active learning strategy was employed to select informative unlabeled images for manual annotation.
• Synthetic images were generated by combining cropped instruments and backgrounds from selected images with blending techniques.
• The framework integrates active learning with synthetic data generation for iterative neural network training.

Main Results:

• The proposed active learning-based synthetic image generation framework demonstrated significant performance improvements in instrument segmentation.
• Effectiveness was validated across multiple surgical datasets, including sinus and intra-abdominal surgeries.
• Performance gains were particularly notable when working with limited amounts of annotated data.

Conclusions:

• The combined approach of active learning and synthetic data generation effectively alleviates the data annotation burden in surgical instrument segmentation.
• This method offers a practical solution for improving the accuracy and efficiency of deep learning models in minimally invasive surgery.
• The open-sourced code facilitates further research and application of this technique.