An attention framework based on multiple color spaces for artifact classification in whole slide images

Area of Science:

• Digital pathology
• Medical imaging analysis
• Artificial intelligence in healthcare

Background:

• Computational pathology models are susceptible to artifacts from tissue preparation and whole-slide imaging.
• Artifacts like folds, uneven staining, and contamination degrade image quality and reduce model reliability.

Purpose of the Study:

• To develop a robust artifact classification method for histopathological images.
• To improve the reliability of computational pathology models by addressing image artifacts.

Main Methods:

• Introduced AC-AttnNet, a novel multicolor-space network for artifact classification.
• Transformed input images into RGB, HSV, HED, and CIELAB color spaces for feature extraction using Attention-Enhanced Residual Blocks (AERBs).
• Employed a Cross-Color Attention Module (CCAM) for dynamic fusion of color-domain representations via attention matrices.

Main Results:

• AC-AttnNet achieved a mean test accuracy of 97.91 ± 0.16% and 99.10% validation accuracy on a breast histopathology artifact dataset.
• Demonstrated strong generalization with 87.3% average accuracy on cross-dataset evaluation, outperforming existing methods.
• Ablation studies confirmed the benefits of multi-color-space modeling and CCAM fusion for improved accuracy.

Conclusions:

• AC-AttnNet effectively classifies artifacts in histopathological images, enhancing computational pathology model reliability.
• The proposed cross-color feature interaction approach is crucial for robust artifact classification.
• AC-AttnNet is a promising tool for quality control and restoration in digital pathology pipelines.