Leveraging Text-Modulated Semantic Guidance for Low-Light Endoscopic Image Enhancement

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Low light conditions in endoscopic imaging degrade image quality, impacting diagnosis and surgical guidance.
• Existing methods struggle to effectively enhance visibility, contrast, and reduce noise in endoscopic images.
• Pretrained models like CLIP show promise for vision tasks, but require adaptation for specific applications like low-light endoscopic image enhancement (LLEIE).

Purpose of the Study:

• To develop a novel text-modulated semantic-aware discriminator (TMSD) for improving low-light endoscopic image enhancement (LLEIE).
• To leverage pretrained CLIP model priors for enhanced semantic understanding in endoscopic imaging.
• To integrate TMSD into existing enhancement baselines to improve visual restoration without additional inference cost.

Main Methods:

• Investigated pretrained CLIP model priors and embedded them into a text-modulated semantic-aware discriminator (TMSD).
• Developed a prompt learning procedure to obtain text and image semantic priors for normal-light endoscopic imaging.
• Utilized a text modulator to synergize text and image priors, employing convolutional modulation and cross-attention for semantic guidance integration.

Main Results:

• TMSD integration improved perceptual quality in seven representative low-light enhancement baselines across five benchmark datasets.
• Significant improvements were observed in cross-domain clinical generalization scenarios.
• Downstream segmentation accuracy was notably enhanced, demonstrating the framework's practical application potential.

Conclusions:

• The proposed TMSD effectively enhances low-light endoscopic images by leveraging CLIP priors and adversarial learning.
• TMSD offers a versatile solution for improving visual quality and diagnostic accuracy in endoscopic imaging.
• The framework demonstrates broad applicability, successfully adapted to metal artifact reduction tasks with no extra inference cost.