Patch Extraction

The foundational algorithm that systematically crops a WSI into a grid of overlapping or non-overlapping patches. A fixed-size window (e.g., 256x256 or 512x512 pixels at 20x or 40x magnification) traverses the tissue region. Overlap (typically 0-50%) ensures that diagnostically relevant objects bisected by a tile boundary are fully captured in at least one patch. This deterministic process converts a sparse, massive image into a dense, uniformly sized tensor batch suitable for GPU processing.

A critical filtering step that discards patches containing only glass slide background, preserving computational resources. Otsu's thresholding in the HSV or LAB color space is commonly used to generate a binary tissue mask. Only patches with a minimum tissue occupancy ratio (e.g., >50% foreground pixels) are retained. This step can reduce the total patch count by 70-90%, dramatically accelerating downstream inference and training.

WSI files are stored as multi-resolution pyramids. Patches are typically extracted from a high-resolution level (Level 0) for cellular detail, but contextual patches can be extracted from lower magnification levels to capture broader tissue architecture. A dual-stream approach feeds high-mag patches into a CNN for nuclear morphology and low-mag patches into a separate encoder for spatial context, fusing the embeddings for a richer representation.

Instead of a blind grid, patches can be centered on regions of interest identified by color deconvolution or a lightweight detector. For example, hematoxylin channel intensity can guide extraction toward cell-dense areas, while eosin channel analysis can target stromal regions. This intelligent sampling strategy ensures balanced representation of different tissue compartments and reduces the dominance of uninformative stroma or necrosis in the training set.

Each extracted patch must retain its absolute spatial coordinates (x, y) within the WSI and its magnification level. This spatial provenance is essential for reconstructing attention heatmaps, mapping predictions back to original tissue morphology, and enabling graph-based models that rely on patch adjacency. A well-structured manifest file (e.g., HDF5 or Parquet) stores patch paths, coordinates, and associated slide-level metadata for efficient data loading.

Two architectural paradigms exist. Pre-extraction saves all patches to disk (e.g., as JPEG or PNG tiles) for deterministic, repeatable training but requires significant storage (terabytes per cohort). On-the-fly extraction reads the WSI file during training and dynamically samples patches, eliminating storage overhead but demanding high-throughput file I/O and introducing a CPU bottleneck. Hybrid approaches cache frequently accessed patches in RAM or fast SSD tiers.

A high-resolution digital scan of an entire glass pathology slide, producing a gigapixel image file that serves as the source data for patch extraction. WSIs are stored in a multi-resolution pyramid format, where the baseline layer contains the full-resolution pixel data. Patch extraction operates on this high-resolution layer to generate manageable input tiles for convolutional neural networks.

The pixel-level classification that delineates tissue regions from the glass background on a WSI. This critical pre-processing step ensures patch extraction only samples informative tissue areas, discarding empty background. Common approaches include:

• Otsu thresholding on grayscale or HSV color space
• U-Net-based semantic segmentation for precise boundary detection
• Adaptive thresholding to handle staining variability Filtering background patches dramatically reduces computational load and prevents the model from learning irrelevant features.

A weakly supervised learning paradigm where a model is trained on labeled bags of instances. In computational pathology, a WSI is treated as a bag, and its extracted patches are the unlabeled instances. Only the slide-level diagnosis label is required during training. MIL aggregation functions—such as attention-based pooling or max pooling—combine patch-level features into a slide-level prediction, making patch extraction the foundational data preparation step for this widely used framework.

A computational technique to standardize color appearance across pathology images, mitigating variability from different staining protocols, scanner models, and laboratory workflows. Applied either before or after patch extraction, common methods include:

• Macenko normalization: Decomposes RGB into stain concentration matrices
• Reinhard normalization: Matches color distribution to a reference template
• CycleGAN-based style transfer: Learns a mapping between stain domains Consistent stain appearance is essential for ensuring extracted patches share a common color distribution, improving model generalization.

An automated quality control step that identifies and excludes image regions containing tissue folds, air bubbles, pen marks, or out-of-focus areas before or during patch extraction. Artifacts can introduce spurious correlations that degrade model performance. Detection methods include:

• Blur detection via Laplacian variance thresholding
• Color anomaly detection to flag pen marks and cauterization artifacts
• Deep learning classifiers trained on artifact vs. clean tissue patches Filtering artifact patches ensures only high-quality tissue regions enter the training pipeline.

A learned, compact numerical vector representation of a pathology image patch, capturing its morphological essence. After patch extraction, each tile is passed through a pre-trained encoder—often a Vision Transformer (ViT) or ResNet—to produce a fixed-dimensional embedding. These embeddings serve as the input to MIL aggregation models, enabling slide-level classification without requiring patch-level annotations. Foundation models like UNI and CTransPath are specifically pre-trained on histology data to generate rich, generalizable patch embeddings.