Computer vision for biological imaging
Pathologist scoring can vary between observers and across time. Whole slide images also contain far more cells and regions than can be reviewed manually at scale. We build computer vision software for image classification, segmentation, detection, and biomarker quantification.
Image analysis
Classification
Assigns a label to an image or region, such as tumor versus non tumor or one tissue subtype versus another.
Segmentation
Identifies the boundaries of cell nuclei or tissue regions so they can be counted and measured.
Detection
Finds specific objects, such as positive cells for a stain, mitotic figures, or other structures.
Quantification
Converts the image analysis into measurements such as cell density, staining intensity, and the proportion of positive cells in a region.
Convolutional neural networks are used across these tasks. Newer architectures are useful where the model needs context beyond a small image patch.
Model limitations
Training data
Training data comes from annotated images, usually marked by a pathologist or biologist. The annotation can be a substantial part of the project, particularly when the task requires cell level labels or adjudication.
Variation
Scanner, staining, specimen preparation, and laboratory practice vary between sources. A model can learn those differences as artefacts. We test against held out material from other sources during validation.
Interpretability
For clinical work, the score needs to be reviewable. A given model may use heatmaps, attention visualisation, region level results, and the measurements behind the score.
Validation
Models are validated against ground truth that was not used for training. The measures depend on the task, including sensitivity and specificity for detection and agreement with expert scoring for quantification.
Training performance is not used as the validation result.
A public example is a digital pathology platform built for a global healthcare partner. It uses image analysis to quantify HER2, Trop2, CD8, and H&E biomarkers, with automated slide scoring and a real time viewer for pathologists.