DCDSN: dual-color domain siamese network for multi-classification of pathological artifacts

Area of Science:

• Digital Pathology
• Medical Image Analysis
• Artificial Intelligence in Medicine

Background:

• Pathological image artifacts compromise diagnostic accuracy.
• Existing artifact detection methods face challenges with variability, high computational costs, and limited color perception, especially in Hematoxylin-Eosin (H&E) stained images.
• Single-color domain reliance hinders distinguishing artifacts from normal tissue.

Purpose of the Study:

• To develop a robust and efficient method for pathological artifact detection.
• To enhance feature representation and color perception in pathological images.
• To reduce the computational burden of artifact detection systems.

Main Methods:

• Proposed a Dual-Color-Domain Siamese Network (DCDSN) leveraging RGB and HSV color domains.
• Employed Siamese network similarity learning to minimize representation discrepancies between color domains.
• Integrated a lightweight MobileViT-XS backbone with transfer learning for efficiency.
• Introduced a Dynamic Snake Convolution-based feature mapper to improve sensitivity to subtle artifact features.

Main Results:

• The DCDSN achieved 90.97% accuracy, outperforming the AR-Classifier baseline.
• Significantly reduced parameter counts (74.39%) and computation (96.55%).
• Demonstrated strong performance with lower resource consumption.

Conclusions:

• The DCDSN effectively detects artifacts in pathological images with high accuracy and efficiency.
• Leveraging dual color domains and a lightweight architecture enhances feature representation and reduces computational cost.
• The proposed method offers a promising solution for improving pathological image quality and diagnostic reliability.