Determining biomarkers from histopathology slide images

What is this patent about?

’041 is related to the field of digital pathology and automated biomarker detection. Specifically, it addresses the technical challenge of identifying and quantifying cancer-related biomarkers, such as PD-L1, tumor-infiltrating lymphocytes, and molecular subtypes, directly from standard hematoxylin and eosin (H&E) stained histopathology images. The invention seeks to replace or supplement time-consuming and resource-intensive immunohistochemistry (IHC) staining and molecular sequencing with efficient, deep-learning-based analysis of digitized tissue slides.

The underlying idea behind ’041 is the use of specialized deep learning architectures to extract molecular and cellular insights from morphological features that are often invisible to the human eye. The system leverages two primary strategies: a multiscale deep learning framework that integrates broad tissue-level context with high-resolution cell segmentation, and a multiple instance learning (MIL) approach that allows the model to learn from slide-level labels (like RNA sequencing data) without requiring exhaustive, pixel-by-pixel manual annotations by pathologists.

The claims of ’041 focus on a computer-implemented method that separates a digital H&E image into a plurality of tiles and applies them to a deep learning framework to predict biomarker presence. Independent claim 1 specifically protects a multiscale approach that combines tissue classification for each tile with a cell segmentation model to identify individual cells, using the intersection of these data points to predict biomarker status. Claims 23, 26, and 27 focus on a single-scale framework utilizing a trained MIL controller to selectively or randomly discard tiles based on inferred class status, effectively filtering the input data to focus the neural network on the most predictive regions of the tissue.

In practice, the invention functions by processing high-resolution whole-slide images through a pipeline that first normalizes the data to account for staining variations across different labs. For multiscale implementations, the system identifies cell boundaries and interiors using a three-class segmentation model, which allows for accurate cell counting even in dense clusters. For molecular biomarkers like Consensus Molecular Subtypes (CMS), the system uses RNA transcriptome counts to cluster training data, allowing the image-based model to learn the visual signatures associated with specific genetic profiles. This enables the generation of a digital overlay report that maps the probability of biomarker presence across the entire specimen.

This approach differs from prior methods by eliminating the need for dense, localized annotations and specialized IHC staining for every patient. By utilizing weakly supervised learning, the system can be trained on vast datasets where only the overall patient outcome or genetic status is known, rather than requiring a pathologist to manually outline every tumor region. Furthermore, the use of a multi-field of view strategy ensures that the model considers both the immediate cellular morphology and the surrounding stromal environment, leading to more robust predictions of how a patient might respond to specific immunotherapies.