Stain normalization for automated whole-slide image classification
Inventors
Ianni, Julianna • Soans, Rajath Elias • Sankarapandian, Sivaramakrishnan
Assignees
Interested in licensing this patent?
MTEC can help explore whether this patent might be available for licensing for your application.
Publication Number
US-11462032-B2
Publication Date
2022-10-04
Expiration Date
Abstract
Techniques for stain normalization image processing for digitized biological tissue images are presented. The techniques include obtaining a digitized biological tissue image; applying to at least a portion of the digitized biological tissue image an at least partially computer implemented convolutional neural network trained using a training corpus including a plurality of pairs of images, where each pair of images of the plurality of pairs of images includes a first image restricted to a lightness axis of a color space and a second image restricted to at least one of: a first color axis of the color space and a second color axis of the color space, such that the applying causes an output image to be produced; and providing the output image.
Core Innovation
The invention is directed to stain normalization image processing for digitized biological tissue images using a computer implemented convolutional neural network trained using a training corpus comprising a plurality of pairs of images. For each pair, a first image is restricted to a lightness axis of a color space, while a second image is restricted to at least one of a first color axis and a second color axis of the same color space. Applying the trained convolutional neural network produces an output image, and the output image is provided.
The invention characterizes the mapping as learning inter-lab stain appearance variations from one lab to another using an adaptation rather than a fixed target. The paired-image training is constructed by restricting the first input to a lightness axis such as Lab L* and restricting the second input to other color axes such as Lab a* and b*, so that the network learns stain-normalization behavior using color-axis restrictions rather than unconstrained color inputs.
The approach can emphasize robustness to scanner and staining variability and can be trained with training pairs derived from a single laboratory. Optional training constraints include restricting target channels to H&E-derived a*/b* content, augmenting inputs with noise types such as hue noise, saturation noise, brightness noise, contrast noise, and intensity noise, and using rotated training pairs. Digitized whole-slide H&E tissue images are normalized and the normalized output can be provided to downstream automated whole-slide image classification for identifying human dermatopathology classes and subdiagnoses.
Claims Coverage
The document includes two independent claims, a method claim and a system claim. Both independent claims cover applying a trained convolutional neural network to produce and provide a stain-normalized output image using a paired-image training corpus with a first image restricted to a lightness axis and a second image restricted to other color axes.
Paired color-axis convolutional neural network stain normalization training
a computer implemented convolutional neural network trained using a training corpus comprising a plurality of pairs of images, wherein each pair comprises a first image restricted to a lightness axis of a color space and a second image restricted to at least one of a first color axis of the color space and a second color axis of the color space, whereby applying causes an output image to be produced
Stain-normalization system with obtaining, applying, and providing output
at least one electronic processor and at least one persistent electronic memory comprising computer readable instructions that configure the at least one electronic processor to perform operations comprising obtaining a digitized biological tissue image; applying to at least a portion the digitized biological tissue image an at least partially computer implemented convolutional neural network trained using the training corpus; and providing the output image
Across both independent claims, the core claim coverage lies in using a paired-image training corpus where the first input is restricted to a lightness axis and the second input is restricted to other color axes, and in applying a trained convolutional neural network to generate and provide a stain-normalized output image for digitized biological tissue images.
Stated Advantages
Robustness to scanner and staining variability
Documented Applications
Providing a normalized output to downstream automated whole-slide image classification for identifying human dermatopathology classes and subdiagnoses
Interested in licensing this patent?