Artifact Detection

Artifact detection acts as a pre-analytical filter that automatically identifies and excludes image regions containing non-diagnostic material before any tissue analysis occurs.

• Operates as a binary classifier on extracted patches: artifact vs. viable tissue
• Prevents garbage-in, garbage-out failures in slide-level classification
• Typically runs as the first step in a pathology pipeline, before tissue segmentation or feature extraction
• Reduces manual quality control review time by 90%+ in high-throughput labs

Detection systems must generalize across a taxonomy of common preparation and scanning defects that appear in whole slide images.

• Tissue folds: Overlapping tissue sections creating dense, hyperchromatic regions
• Air bubbles: Circular voids introduced during coverslipping, appearing as bright empty spaces
• Pen marks: Deliberate annotations from pathologists that can confuse diagnostic models
• Blur and out-of-focus regions: Scanner autofocus failures producing non-diagnostic patches
• Dust and debris: Particulate matter on the slide or scanner optics

Building robust artifact detectors requires carefully curated datasets that represent the full diversity of real-world slide preparation artifacts.

• Requires pixel-level annotations or patch-level labels from trained histotechnicians
• Must span multiple scanner vendors (Leica, Hamamatsu, Philips) to avoid hardware-specific overfitting
• Class imbalance is severe: artifacts typically represent <5% of total tissue area
• Focal loss is commonly used to address this imbalance during training
• Synthetic artifact generation via data augmentation can supplement scarce real examples

Artifact detection directly feeds into Multiple Instance Learning workflows by filtering the bag of patches before aggregation.

• Artifact patches are excluded from the bag rather than assigned a zero weight
• Prevents attention mechanisms from erroneously focusing on pen marks as diagnostic features
• Reduces computational load by eliminating 30-50% of non-informative patches
• Must operate at inference speed compatible with real-time slide analysis (<2 minutes per gigapixel WSI)

Artifact detection performance is measured using metrics that account for the severe class imbalance inherent in quality control tasks.

• Precision: Critical to avoid discarding viable diagnostic tissue (false positives)
• Recall: Essential to catch all artifacts that could corrupt downstream analysis (false negatives)
• F1-score at the patch level is the standard aggregate metric
• Slide-level pass/fail accuracy: Whether the system correctly flags a slide as requiring re-scanning
• Area under the Precision-Recall curve is preferred over ROC-AUC for imbalanced data

Artifact detectors must maintain performance across unseen domains—new laboratories, staining protocols, and scanner models not represented in training data.

• Stain color and intensity vary dramatically between institutions
• Stain normalization is often applied as a pre-processing step before artifact classification
• Self-supervised pre-training on diverse histology data improves out-of-distribution robustness
• Continuous monitoring for concept drift is necessary as new artifact types emerge in production

The pixel-level classification that delineates tissue regions from the glass background on a whole slide image. This critical pre-processing step identifies the foreground area of interest before artifact detection can proceed. Segmentation masks are generated using thresholding on HSV color space or lightweight convolutional neural networks, producing binary maps that exclude empty slide areas and direct downstream analysis to tissue-containing regions only.

A computational technique to standardize color appearance across pathology images, mitigating variability from different staining protocols, scanner models, and laboratory procedures. Without normalization, a model trained at one institution may fail at another. Common approaches include Macenko's method for stain matrix estimation and cycle-consistent GANs for style transfer. Proper normalization ensures artifact detection algorithms generalize across sites rather than learning spurious color correlations.

The process of tessellating a gigapixel whole slide image into smaller, manageable tiles for processing by convolutional neural networks. Typical patch sizes range from 256×256 to 1024×1024 pixels at 20x or 40x magnification. Artifact detection often operates at the patch level, classifying each tile as usable or rejected. Efficient extraction requires handling irregular tissue boundaries and overlapping sliding windows to ensure complete coverage without redundant computation.

Techniques applied to pathology patches to artificially expand training dataset diversity and improve model robustness. For artifact detection, augmentations include:

• Rotation and flipping to handle variable tissue orientation
• Color jittering to simulate staining variation
• Elastic deformations to mimic tissue folding artifacts
• Cutout and mixing to train models on occluded regions These strategies prevent overfitting to specific artifact appearances present in limited training data.

The automated workflow that filters out non-diagnostic tissue regions before analysis. A typical pipeline executes sequentially:

1. Tissue segmentation to locate foreground
2. Blur detection using Laplacian variance to reject out-of-focus regions
3. Artifact classification to identify folds, bubbles, and pen marks
4. Coverage analysis to ensure sufficient tissue remains after filtering Only patches passing all quality gates proceed to diagnostic model inference, ensuring reliable slide-level predictions.

A high-resolution digital scan of an entire glass pathology slide, producing a gigapixel image file typically stored in a pyramidal, multi-resolution format. Standard WSI scanners capture images at 40x magnification (0.25 microns per pixel) , generating files exceeding 100,000 × 100,000 pixels. Artifact detection must operate efficiently at this scale, often using coarse-to-fine strategies that first screen at low resolution before detailed analysis of suspicious regions at higher magnifications.