Leveraging advanced feature extraction for improved kidney biopsy segmentation.

Area of Science:

• Medical Image Analysis
• Computational Pathology
• Digital Health

Background:

• Renal histopathology segmentation is challenging due to complex glomerular morphology and low contrast.
• Existing models like Segment Anything Model (SAM) struggle with fine details, computational load on whole-slide images (WSIs), and domain-specific features.

Purpose of the Study:

• To develop an enhanced framework, V-SAM, that adapts foundation models for precise medical image segmentation.
• To improve the segmentation of glomeruli in renal histopathology for better disease diagnosis and biomarker discovery.

Main Methods:

• Introduced V-SAM, a novel framework integrating a V-shaped adapter with multi-scale skip connections to preserve spatial hierarchies.
• Employed lightweight adapter layers for efficient fine-tuning of SAM's encoder on histopathology textures.
• Implemented a dynamic point-prompt mechanism for sub-pixel refinement of glomerular boundaries using gradient-aware localization.

Main Results:

• V-SAM achieved state-of-the-art performance on the HuBMAP datasets.
• Demonstrated high accuracy (surpassing 89.31% and 97.65%) and F1-scores (surpassing 86.17% and 95.54%).
• Successfully recovered capillary-level details often lost in standard segmentation models.

Conclusions:

• V-SAM offers a scalable solution for adapting foundation models to clinical workflows in resource-constrained settings.
• This approach bridges the gap between generalizability of foundation models and the precision required for medical imaging.
• V-SAM has direct applications in chronic kidney disease diagnosis and biomarker discovery.