Patch Extraction

The systematic subdivision of a gigapixel image into a grid of smaller image patches (e.g., 256x256 or 512x512 pixels). The stride (step size) between patches determines the degree of overlap. Overlapping tiles (e.g., 50% overlap) ensure that objects bisected by a tile boundary are fully captured in an adjacent tile, preventing information loss at edges. Non-overlapping tiling maximizes computational efficiency but risks missing boundary features.

A critical filtering step that discards patches containing only glass background (white space). An Otsu thresholding algorithm is typically applied to a low-resolution thumbnail to generate a binary tissue mask. Only patches overlapping with the tissue region by a minimum percentage (e.g., >60%) are extracted for inference. This reduces the computational load by 70-90% by ignoring empty space.

Patches are not always extracted from the highest magnification (Level 0). A gigapixel pyramid allows extraction at different downsampled levels. Context patches are often extracted at a lower magnification (e.g., 5x) to capture tissue architecture, while detail patches are extracted at high magnification (e.g., 40x) to analyze cellular morphology. Multi-scale models fuse these views for robust classification.

Every extracted patch must retain its spatial provenance. The filename or a database record stores the original (x, y) coordinates and the magnification level. This spatial indexing is essential for downstream tasks like heatmap generation, where per-patch predictions are stitched back into a probability map aligned with the original WSI geometry.

Raw pixel values are normalized to a standard distribution (e.g., [0, 1] or mean=0, std=1) using ImageNet statistics or dataset-specific parameters. Stain normalization (e.g., Macenko or Vahadane methods) is often applied during or immediately after extraction to correct color variability between different laboratory staining protocols, ensuring the model sees consistent input.

Extracted patches are typically stored as PNG or JPEG files, or serialized into binary formats like TFRecords or HDF5 for high-throughput training. Lossless compression (PNG) preserves fine cellular detail, while lossy compression (JPEG) reduces storage footprint. The choice balances I/O throughput against diagnostic fidelity, with most clinical pipelines opting for lossless storage.

The pixel-level classification that distinguishes tissue regions from glass background before patch extraction begins. Without tissue segmentation, a naive sliding window approach would waste 70-90% of computational resources processing empty background tiles. Segmentation algorithms typically operate on low-resolution thumbnail images to generate a binary tissue mask, which then guides selective patch sampling. Common approaches include Otsu thresholding on HSV color space, adaptive thresholding, and deep learning-based semantic segmentation using architectures like U-Net.

A multi-resolution storage structure that stores a WSI as a sequence of downsampled images, each half the resolution of the previous level. The pyramid enables efficient pan-and-zoom navigation without loading the full-resolution base layer. Patch extraction typically targets a specific pyramid level—level 0 for maximum detail at 40x magnification, or higher levels for lower-resolution context. Understanding pyramid level indexing is critical: extracting a 256x256 patch at level 2 from a 40x scan yields an effective field of view equivalent to 1024x1024 pixels at full resolution.

A color standardization process applied to extracted patches to reduce variability introduced by different laboratory staining protocols, scanner models, and reagent batches. Without normalization, a model trained on patches from Hospital A may fail on patches from Hospital B due to color distribution shift. Techniques include Macenko normalization for Hematoxylin and Eosin decomposition, Reinhard normalization for matching target statistics in LAB color space, and CycleGAN-based generative approaches. Normalization can be applied pre-extraction on the WSI or post-extraction on individual patches.

A weakly supervised learning paradigm that consumes extracted patches as a 'bag of instances' to produce a single slide-level diagnosis. In the MIL framework, a WSI is a labeled bag, and its extracted patches are unlabeled instances. The model learns to aggregate patch-level features—often using attention-based pooling—to classify the entire slide without requiring exhaustive pixel-level annotations. This paradigm directly determines patch extraction strategy: random sampling, hard negative mining, or top-K attention selection all influence which patches are fed to the model during training.