EndoSRR: a comprehensive multi-stage approach for endoscopic specular reflection removal

Area of Science:

• Medical Imaging
• Computer Vision
• Surgical Technology

Background:

• Specular reflections in endoscopic images degrade visual quality.
• These reflections hinder the performance of computer vision algorithms.
• Existing methods face challenges due to reflection variability and limited datasets.

Purpose of the Study:

• To develop a robust method for eliminating specular reflections in endoscopic images.
• To enhance visual perception and computer vision algorithm performance for downstream tasks.
• To improve the accuracy and safety of minimally invasive surgery.

Main Methods:

• EndoSRR employs a two-stage approach: reflection detection and inpainting.
• Reflection detection utilizes a fine-tuned Segment Anything Model (SAM) for accurate mask generation.
• Reflection region inpainting uses LaMa, a Fourier convolution-based model, refined by a dual pre-trained model iterative optimization strategy (DPMIO).

Main Results:

• EndoSRR outperforms state-of-the-art methods on the SCARED-2019 dataset.
• Qualitative results show accurate reflection detection and natural inpainting.
• Quantitative evaluations demonstrate superior performance in segmentation (IoU, E-measure) and inpainting (PSNR, SSIM) metrics.

Conclusions:

• Proficient endoscopic specular reflection removal is crucial for enhancing visual perception and downstream tasks.
• The proposed methodology and results are expected to advance specular reflection removal techniques.
• This advancement will contribute to improved accuracy and safety in minimally invasive surgery.