Whole-Slide Image (WSI)

A WSI is stored as a pyramidal data structure containing multiple downsampled versions of the base image. The highest resolution (Level 0) represents the native scanner magnification, often 0.25 microns per pixel at 40x. Successive levels are generated by halving dimensions, enabling efficient zooming and panning without loading the entire gigapixel image into memory. This architecture is critical for computational pathology pipelines, where algorithms typically operate on extracted patches at a specific pyramid level.

To manage massive file sizes, WSIs are internally organized as a grid of small image tiles, typically 256x256 or 512x512 pixels. Access is mediated by specialized libraries such as OpenSlide and Bio-Formats, which provide a unified API for reading tiles from diverse proprietary file formats (e.g., .svs, .ndpi, .mrxs). This tiled architecture enables random access to any image region without decoding the entire file, a fundamental requirement for training deep learning models on millions of extracted patches.

WSIs exhibit significant color and intensity variation due to differences in staining protocols, scanner models, and reagent batches. This domain shift poses a major challenge for model generalization. Stain normalization techniques, such as Macenko or Vahadane methods, decompose images into stain density maps to standardize appearance. Without robust color augmentation during training, a model developed on one scanner may fail catastrophically on images from another institution.

Beyond pixel data, WSI files embed rich metadata including objective magnification, scanner type, and physical pixel spacing. Some formats also store associated images such as a macro overview, label image, and thumbnail. In clinical workflows, the WSI is linked to the DICOM standard via the VL Whole Slide Microscopy Image IOD, which encapsulates the pixel pyramid alongside structured patient, specimen, and staining metadata, enabling integration with PACS and EHR systems.

WSIs frequently contain artifacts that can mislead diagnostic algorithms:

• Tissue folds: Dark linear regions caused by sectioning wrinkles
• Pen marks: Surgical inking that can be mistaken for positive staining
• Air bubbles: Circular voids introduced during coverslipping
• Out-of-focus regions: Blur due to uneven tissue section thickness Automated image quality control pipelines must detect and exclude these regions before analysis to prevent false positive biomarker quantification.

WSIs employ lossy and lossless compression schemes, primarily JPEG and JPEG 2000, to reduce storage footprints from tens of gigabytes to manageable sizes. JPEG 2000 is preferred for its support for progressive decoding and region-of-interest access. However, compression artifacts at high ratios can introduce high-frequency noise that degrades the performance of texture-sensitive algorithms. Understanding the compression level is critical for reproducibility in computational pathology research.

The interdisciplinary field applying machine learning and image analysis algorithms to digitized tissue slides for automated diagnosis, biomarker discovery, and outcome prediction. It transforms qualitative visual assessment into quantitative, reproducible data.

• Replaces manual microscopy with high-throughput computation
• Enables discovery of novel morphological biomarkers invisible to the human eye
• Bridges pathology with bioinformatics and oncology informatics

A weakly-supervised learning paradigm where a model is trained using only slide-level labels by aggregating predictions from unlabeled patches extracted from a gigapixel WSI. The slide is a 'bag' and patches are 'instances'.

• Eliminates the need for costly pixel-level annotation
• Uses attention-based pooling to identify diagnostically relevant regions
• Standard approach for cancer classification and survival prediction

A computational pre-processing technique that standardizes the color appearance of histology images to reduce variability caused by different staining protocols, scanner models, and laboratory reagents.

• Aligns color distributions to a reference template
• Critical for generalizable deep learning models
• Methods include Macenko, Vahadane, and cycle-GAN approaches

A deep learning task that classifies every pixel in a WSI into predefined tissue categories such as tumor, stroma, necrosis, or normal epithelium. It provides spatial context without distinguishing individual object instances.

• Architectures: U-Net, DeepLab, SegFormer
• Enables tumor-stroma ratio quantification
• Foundation for tissue composition analysis

The high-throughput extraction and mining of hundreds of quantitative morphological, textural, and spatial features from digital pathology images to characterize tumor heterogeneity.

• Extracts features like nuclear shape, chromatin texture, and glandular architecture
• Combines with radiomics for multi-modal analysis
• Feeds into prognostic model development

A neural architecture that applies the self-attention mechanism to sequences of image patches, capturing long-range spatial dependencies for state-of-the-art pathology image classification.

• Processes WSI as sequences of 16x16 pixel patches
• Captures global tissue architecture better than CNNs
• Powers foundation models like UNI and Virchow