Computational Pathology

The foundational step converting glass histology slides into gigapixel digital images using specialized scanners. A single Whole Slide Image (WSI) can exceed 100,000 x 100,000 pixels at 40x magnification, requiring multi-resolution pyramidal storage formats. Key considerations include scanner calibration, focus quality, and compression artifacts that downstream algorithms must accommodate.

The process of tessellating a gigapixel WSI into thousands of smaller, manageable image tiles (patches) for processing by convolutional or transformer-based neural networks. Typical patch sizes range from 224x224 to 1024x1024 pixels. Strategies include:

• Grid-based extraction with fixed stride
• Tissue-informed sampling that skips background regions
• Random sampling for efficient training Patch-level analysis is the gateway to both supervised and weakly supervised learning paradigms.

A computational technique to standardize color appearance across pathology images from different laboratories. Variability arises from differences in:

• Staining protocols and reagent batches
• Scanner models and color calibration
• Tissue preparation methods Algorithms like Macenko, Vahadane, and Reinhard normalization map source images to a reference color distribution. This is essential for model generalization, preventing spurious correlations with stain intensity rather than morphological features.

The transformation of each image patch into a compact numerical vector (embedding) that captures its morphological essence. Modern approaches leverage:

• Self-Supervised Learning (SSL) pre-trained on millions of unlabeled histology patches
• Vision Transformers (ViT) applying self-attention across patch sequences
• Foundation models serving as general-purpose feature extractors These embeddings encode cellular morphology, tissue architecture, and textural patterns into a latent space where semantically similar structures cluster together, enabling efficient downstream aggregation.

A gigapixel digital scan of an entire glass pathology slide, forming the raw data substrate for computational analysis. WSIs are multi-resolution pyramidal files where patch extraction tessellates the image into manageable tiles for convolutional neural networks. The sheer scale—often exceeding 100,000×100,000 pixels—demands specialized streaming protocols and tissue segmentation to separate foreground from glass background before any model inference begins.

A weakly supervised learning paradigm where a model is trained on labeled bags of instances rather than individually annotated examples. In pathology, a WSI is a bag containing thousands of unlabeled patches; only the slide-level diagnosis is known. Attention-based MIL dynamically weights diagnostically relevant patches, enabling slide-level classification without exhaustive pixel-level annotations. Frameworks like CLAM (Clustering-constrained Attention MIL) exemplify this approach.

A computational pre-processing technique that standardizes color appearance across pathology images to mitigate variability introduced by:

• Different staining protocols and reagent batches
• Scanner hardware variations
• Tissue preparation differences Methods range from histogram matching to deep learning-based style transfer using generative adversarial networks. Without normalization, a model may learn spurious color correlations rather than morphological features, catastrophically degrading domain generalization.

A standardized histological grading system for prostate cancer based on architectural glandular patterns, scored from 1 (well-differentiated) to 5 (poorly differentiated). Modern deep learning models automate Gleason scoring directly from H&E-stained WSIs, assigning both a primary and secondary pattern to produce a Gleason sum. This task exemplifies fine-grained image classification requiring models to distinguish subtle textural and structural differences in tissue architecture.

Immune cells that have migrated into the tumor microenvironment, quantified as a prognostic and predictive biomarker. Computational pathology pipelines automatically segment and count TILs from H&E-stained slides using semantic segmentation models. High TIL density often correlates with favorable immunotherapy response, making automated quantification critical for immuno-oncology clinical trials and treatment planning.

A large-scale AI model pre-trained on massive, diverse histology datasets using self-supervised learning (SSL). These models serve as general-purpose feature extractors, generating rich feature embeddings for downstream tasks like cancer subtyping or biomarker prediction. By learning universal visual representations from millions of unlabeled patches, foundation models dramatically reduce the annotated data required for specific diagnostic applications, embodying the pre-train then adapt paradigm.