Exploiting the potential of unlabeled endoscopic video data with self-supervised learning.

Area of Science:

• Surgical data science
• Medical image analysis
• Deep learning in medicine

Background:

• Surgical data science aims to optimize patient treatment through comprehensive observation.
• Deep learning advancements in automatic image annotation highlight a bottleneck in acquiring labeled data for algorithm training.
• Self-supervised learning offers a potential solution to the data annotation challenge.

Purpose of the Study:

• To investigate the efficacy of self-supervised learning for training models in surgical data science.
• To address the bottleneck of limited reference annotations for algorithm development.
• To explore methods for leveraging unlabeled data to improve model pre-training.

Main Methods:

• Utilized unlabeled endoscopic video data for pre-training convolutional neural networks (CNNs).
• Proposed a re-colorization task using a conditional generative adversarial network (cGAN) as an auxiliary task for initialization.
• Investigated a variant with a second pre-training step using labeled data from a related domain.
• Validated the approach using medical instrument segmentation as the target task.

Main Results:

• The self-supervised approach significantly reduced manual annotation effort, decreasing the need for labeled images by up to 75% compared to training from scratch.
• Achieved comparable or superior performance to baseline methods without sacrificing accuracy.
• Outperformed alternative pre-training strategies, including using public medical and non-medical datasets.

Conclusions:

• The proposed self-supervised learning method efficiently utilizes available labeled and unlabeled data.
• This approach has the potential to become a valuable tool for pre-training CNNs in surgical data science.
• It offers a scalable solution to the data annotation challenge in medical AI development.