Pathologists do not always score the same slide the same way
The same pathologist on two different days may not agree with themselves. Biological images carry information that is slow to extract by eye and inconsistent between observers. Computer vision brings consistency to that scoring, and once a model is validated it can process volumes no lab could staff for.
What the models do
Most biological imaging work reduces to a few core tasks.
Classification
Assigns a label to an image or region — tumor or not, this subtype or that one.
Segmentation
Draws boundaries, outlining every cell nucleus or every region of tissue type, so they can be counted and measured.
Detection
Finds specific objects — positive cells for a given stain, mitotic figures, particular structures.
Quantification
Turns those outputs into numbers a scientist can use: cell density, staining intensity, the proportion of positive cells in a region.
Convolutional neural networks handle most of this, with newer architectures gaining ground for tasks that need context across a whole slide rather than a small patch.
Why it is harder than it looks
Training data
A model learns from annotated examples, and annotation in this field means a pathologist or biologist marking up images by hand. That is slow, expensive, and the bottleneck for almost every project.
Variation
Slides from different labs, different scanners, different staining batches look different in ways a model trained on one source may not generalize across. Stain normalization and careful validation on held out data from other sources are the standard defenses, and skipping them is how a model that looked excellent in development fails in production.
Interpretability
A model that outputs a score with no explanation is hard for a clinician to trust and hard for a regulator to accept. Heatmaps, attention visualization, and reporting which regions drove a decision are part of the engineering, not an optional extra.
Validation
A model intended for clinical or research use has to be validated against ground truth it has never seen, with performance measured on the metrics that matter for the task — sensitivity and specificity for detection, agreement with expert scoring for quantification.
Performance on training data is not evidence of anything.
The most concrete public example is a digital pathology platform built for a global healthcare partner, using image analysis to quantify biomarkers including HER2, Trop2, CD8, and H&E, with automated slide scoring and a real time viewer for pathologists.