Pathology AI 101

Whole-slide images are gigapixel pyramids. This lesson is about getting comfortable moving through one before any AI is involved.

Two ways to find nuclei: classic colour-intensity thresholding, and a deep-learning model (StarDist) that learned what a nucleus looks like.

Once nuclei are segmented, each one becomes a row of measurements. This lesson covers what those measurements are and what they mean.

Features are only useful if they separate cells you care about. Here they are used to pick out tumour cells from everything else.

Mitotic figures are rare and easy to confuse with apoptotic cells. Active learning lets the model ask for the labels it needs most.

Moving from cells to regions. Segmenting a slide into tumour, stroma, necrosis and fat, first with colour rules and then with a deep network.

Interactive segmentation turns a few clicks or a rough scribble into a full region. VistaPath is the tool in TissueLab that does this on pathology slides.

A tissue mask is the basis for numbers: how much tumour, how much stroma, and what the cells inside each region look like.

A model always produces an answer. This lesson is about telling a good answer from a bad one, and what a bad one costs downstream.

When only the slide has a label, multiple instance learning (MIL) learns which patches explain it. Built here on ResNet-18 patch embeddings.

Vision-only foundation models such as UNI were pre-trained on millions of pathology patches. Their embeddings replace ResNet-18 and usually beat it.

Slide-level encoders take thousands of patch embeddings and produce one vector for the whole slide, ready for downstream tasks.

Putting the slide-level toolkit to work on the three questions clinicians ask most: how long, will it respond, and which kind.

Immunohistochemistry (IHC) shows where a protein is. Turning brown stain into a score means detecting cells, separating stain, and counting.

iSight is a workflow for separating patients who respond to a therapy from those who do not, directly from their slides.

Spatial transcriptomics is expensive. Models trained on paired data can predict gene expression at each spot from the H&E image alone.

Extending the same idea to several modalities at once: transcriptome, proteome and epigenome, predicted jointly from morphology.

Models trained on image-text pairs put a patch and a sentence in the same space, so a slide can be searched with words.

PathChat is a LLaVA-style assistant: a pathology vision encoder wired into a large language model, so you can show it a region and ask about it in plain language.

An agent takes a goal in plain language, chooses tools, runs them, and checks its own work. This lesson shows what that looks like on a slide.

Agents in TissueLab do not stay fixed. Feedback, new tools and self-evolution change what they can do over time.