Image Quality Control

Identifies regions where tissue sections have folded over during slide preparation, creating dense, hyperchromatic areas that obscure cellular morphology. Convolutional neural networks analyze texture patterns and sharp intensity transitions characteristic of folds.

• Detects sharp gradient transitions at fold boundaries
• Differentiates folds from genuine dense cellular regions
• Excludes folded areas from downstream analysis masks

Locates and masks surgical inking and pathologist annotations that appear as high-intensity, saturated regions on digitized slides. These foreign ink deposits can be misinterpreted as positive staining in IHC analysis.

• Color deconvolution separates ink from biological stains
• Spectral analysis distinguishes pen pigments from tissue chromogens
• Prevents false-positive biomarker quantification

Detects circular, out-of-focus artifacts caused by trapped air during coverslipping. Morphological operations combined with focus quality metrics identify these regions that lack tissue information entirely.

• Circularity and edge characteristics distinguish bubbles from luminal spaces
• Excluded from tissue area calculations to prevent dilution bias
• Critical for accurate tumor-stroma ratio computation

Quantifies image sharpness across the slide to flag out-of-focus regions where diagnostic interpretation is unreliable. Laplacian variance and frequency-domain analysis measure local blur severity.

• Patch-level focus scores generate a quality heatmap
• Threshold-based rejection of regions below diagnostic sharpness
• Prevents AI models from learning from degraded visual features

Segments actual tissue regions from the glass background to constrain all subsequent analysis to biologically relevant areas. Otsu thresholding in HSV color space robustly separates tissue from empty slide background.

• Reduces computational load by ignoring background pixels
• Prevents background noise from contaminating feature extraction
• Essential first step in any WSI processing pipeline

Systematically identifies and excludes image patches that fail to meet minimum sharpness thresholds due to z-stack acquisition errors or tissue thickness variation. Ensures only diagnostically interpretable content reaches pathologists and AI models.

• Brenner gradient and Tenengrad algorithms quantify local focus
• Adaptive thresholding accounts for tissue-dependent focus expectations
• Integrated into real-time WSI scanning quality feedback loops

A systematic classification of non-diagnostic image content that must be excluded before analysis:

• Tissue folds: Overlapping sections causing dense, hyperchromatic linear regions
• Pen marks: Surgical inking delineating margins, appearing as dense blue/black regions
• Air bubbles: Circular, translucent voids introduced during coverslipping
• Blur/out-of-focus: Regions lacking high-frequency edge information due to z-stack misalignment
• Dust/debris: Particulate matter on the slide or scanner optics
• Uneven illumination: Vignetting or shading artifacts from the scanner light path

Modern artifact detection employs convolutional neural networks trained on pixel-level annotations to generate binary exclusion masks.

The initial step in quality control that separates tissue regions from the glass background. This process:

• Applies Otsu thresholding on the saturation channel of HSV color space
• Generates a binary mask identifying foreground pixels
• Removes white-space background to reduce computational load
• Enables patch-based sampling only from diagnostically relevant areas

Advanced methods use U-Net architectures for robust segmentation across varied staining protocols. The resulting tissue mask gates all downstream analysis, ensuring models process only biological material.

A computational metric that quantifies image sharpness to reject blurred regions. Key algorithms include:

• Laplacian variance: Measures the spread of second-derivative pixel intensities; low variance indicates blur
• Brenner gradient: Computes squared differences between pixels separated by a fixed offset
• Tenengrad function: Aggregates gradient magnitudes above a Sobel threshold
• Frequency domain analysis: Evaluates power spectral density decay rates

A focus score is calculated per tile; regions falling below an empirically determined threshold are excluded. This is critical for nuclear segmentation tasks where precise membrane boundaries are essential.

A pre-processing technique that standardizes color appearance across slides to mitigate inter-laboratory variability. Methods include:

• Reinhard normalization: Matches mean and standard deviation in LAB color space to a reference image
• Macenko method: Decomposes RGB into stain-specific optical density vectors using singular value decomposition
• Structure-preserving color normalization (SPCN): Leverages sparse non-negative matrix factorization
• CycleGAN-based approaches: Learn unpaired image-to-image translation for stain transfer

Normalization is essential before quality control to ensure artifact detectors generalize across staining protocols and scanner types.

A granular approach where each extracted tile receives a composite quality score aggregating multiple criteria:

• Tissue percentage: Ratio of foreground to background pixels
• Artifact overlap: Intersection-over-union with artifact masks
• Focus metric: Normalized sharpness value
• Stain intensity: Deviation from expected hematoxylin/eosin optical density ranges

Tiles are ranked and filtered using configurable thresholds. This quality-aware sampling ensures that only diagnostically viable patches enter the Multiple Instance Learning (MIL) pipeline, directly impacting model performance.

Slide-level aggregate statistics reported to laboratory information systems for continuous quality improvement:

• Artifact burden: Percentage of total tissue area affected by artifacts
• Coverage ratio: Proportion of the slide containing tissue vs. blank glass
• Focus uniformity: Coefficient of variation of focus scores across the slide
• Stain quality index: Deviation from reference hematoxylin/eosin color distributions

These metrics enable real-time feedback to histology technicians, flagging slides requiring re-scanning or re-staining before pathologist review. Integration with LIS/LIMS systems automates quality assurance workflows.