Detecting cells of interest in large image datasets using artificial intelligence

Inventors

WILKINSON, Jr., Dan Charles • BURNETT, Benjamin Adam

Assignees

BlueRock Therapeutics LP

Abstract

A method for selecting a final model for detecting cells of interest in image datasets includes dividing a curated image dataset into a training set, a validation set, and a testing set where each image in the curated image dataset has been labeled as positive or negative for a cell of interest. The method trains each model of an ensemble of neural networks using the training and validation sets. Next, each model of the ensemble is tested using the testing set and the predictions of the ensemble are combined. The combined prediction is compared to the label and the method determines whether the combined prediction satisfies a pre-determined level of detection (LOD). If so, the method outputs the ensemble as a final ensemble. If not, the method modifies a hyperparameter of at least one of the models of the ensemble until the combined prediction satisfies the pre-determined LOD.

Core Innovation

The invention provides a method for detecting one or more cells of interest in image datasets by processing stained histological images that each contain an independent channel. Pixel values of the independent channel are binarized, connected pixel areas in the binarized images are identified to determine an area of interest, and each area of interest is defined by bounding coordinates and cropped to generate sub-images for downstream detection.

Each cropped sub-image is labeled as positive or negative for a cell of interest. A final classification model for detecting the cell of interest in each sub-image is selected and deployed to detect cells of interest in unclassified image datasets. The selection supports detection-level evaluation and includes using an ensemble classification model in the family context.

Binarizing of independent channel pixel values includes binarizing pixel values in a set of merged images generated by merging at least the plurality of stained histological images with a first independent channel applied and corresponding stained histological images with a second independent channel applied. In related implementations, an image autoencoder is trained to learn background signal from a set of images lacking cells of interest and is used to remove background noise prior to binarization.

Claims Coverage

The independent claims are directed to a cell-detection pipeline that binarizes independent channels, defines areas of interest via connected pixel areas with bounding coordinates, crops labeled sub-images, selects a final classification model, and deploys it to unclassified image datasets. The independent claims additionally refine how binarization is performed for merged images across fluorescent channels, with further dependent refinements focusing on detection-level model selection, bounding coordinate extrema, and optional background-noise removal by an image autoencoder.

Across the independent claims, the core claim coverage is directed to identifying connected pixel areas in binarized independent-channel stained histological images and cropping them using bounding coordinates, labeling cropped sub-images as positive or negative and training/selecting a final classification model, and deploying the model to detect cells of interest in unclassified datasets. Additional refinements in the family cover merged-image binarization and include dependent features that further specialize bounding coordinates and optional background-noise removal using an image autoencoder prior to binarization.

Stated Advantages

Documented Applications