Whole Slide Image (WSI)

A WSI is stored as a multi-resolution pyramid, not a flat file. The base layer holds the highest resolution (e.g., 0.25 microns per pixel), while successive downsampled layers enable smooth pan-and-zoom navigation. This structure mimics the experience of adjusting microscope magnification, allowing algorithms to efficiently switch between coarse tissue context and fine cellular detail without loading the entire gigapixel image into memory.

While Brightfield H&E is the standard, a WSI can encode far more than RGB. Advanced scanners capture multispectral or fluorescent channels (e.g., DAPI, FITC, Cy5) for multiplexed immunohistochemistry (IHC). Each channel represents a distinct biomarker, and computational deconvolution algorithms can mathematically separate overlapping stain spectra to quantify individual protein expressions on a per-pixel basis.

Unlike a static 2D photo, many scanners capture Z-stacks—multiple focal planes at varying depths through the tissue section. This volumetric data is crucial for cytology specimens where cells float in 3D space. Extended Depth of Field (EDF) algorithms then fuse the sharpest pixels from each plane into a single, all-in-focus 2D composite, ensuring no diagnostic information is lost due to topographical unevenness.

A WSI file encapsulates a vast array of structured metadata beyond pixel data. This includes scanner calibration profiles, objective lens power, micron-per-pixel resolution, and often a label image—a separate macro scan of the slide's physical barcode. In formats like OME-TIFF, this extends to XML-encoded experimental conditions, linking the image directly to the laboratory information system (LIS) for full traceability.

Internally, the pyramid layers are subdivided into discrete rectangular tiles (commonly 256x256 or 512x512 pixels). This tiling is the critical enabler for parallel processing; a Patch Extraction engine can fetch thousands of non-contiguous tiles simultaneously across a GPU cluster without decoding the entire image. This architecture transforms a massive file into a random-access database of visual tokens.

WSI data exists in a fragmented ecosystem of vendor-specific formats (e.g., .svs, .ndpi, .mrxs) that often wrap TIFF structures with custom compression. Interoperability is achieved through libraries like OpenSlide, which abstract these proprietary headers into a unified API. The push toward OME-TIFF and DICOM Supplement 145 aims to standardize this, embedding clinical context directly into the pixel data for true enterprise portability.

The process of dividing a massive whole slide image into smaller, manageable image tiles that can be processed by a convolutional neural network. Typical patch sizes range from 256x256 to 1024x1024 pixels at 20x or 40x magnification. Extraction logic must filter out background glass and adipose tissue to avoid wasting compute on non-diagnostic regions, often using Otsu thresholding on a low-resolution thumbnail.

A weakly supervised learning paradigm where a model is trained on labeled bags of instances, making it ideal for slide-level classification from patch-level data without exhaustive annotations. In WSI analysis, the entire slide is the bag and extracted patches are the instances. Only the slide-level diagnosis is required for training, eliminating the bottleneck of pixel-level annotation.

A computational process that standardizes the color appearance of histological images to reduce variability between different laboratory staining protocols and scanner hardware. Techniques include Macenko color deconvolution, Reinhard normalization, and cycle-GAN-based approaches. Without normalization, a model may learn spurious color correlations rather than morphological features.

The process of rendering a color-coded probability overlay on a whole slide image to visualize the spatial distribution of model predictions or regions of high diagnostic interest. Heatmaps map per-patch inference scores back to their original spatial coordinates, allowing pathologists to visually verify which tissue regions drove the model's classification decision.