Systems and methods for estimating histological features from medical images using a trained model
Inventors
Assignees
Publication Number
US-12171542-B2
Publication Date
2024-12-24
Expiration Date
2037-03-09
Interested in licensing this patent?
MTEC can help explore whether this patent might be available for licensing for your application.
Abstract
Systems and methods for estimating quantitative histological features of a subject's tissue based on medical images of the subject are provided. For instance, quantitative histological features of a tissue are estimated by comparing medical images of the subject to a trained model that relates histological features to multiple different medical image contrast types, whether from one medical imaging modality or multiple different medical imaging modalities. In general, the trained model is generated based on medical images of ex vivo samples, in vitro samples, in vivo samples or combinations thereof, and is based on histological features extracted from those samples. A machine learning algorithm, or other suitable learning algorithm, is used to generate the trained model. The trained model is not patient-specific and thus, once generated, can be applied to any number of different individual subjects.
Core Innovation
The invention provides systems and methods for estimating quantitative histological features of a subject’s tissue using medical images. This is accomplished by comparing medical images to a trained model that relates histological features to various medical image contrast types, either from one imaging modality or multiple modalities. The trained model is created based on medical images and histological features extracted from ex vivo, in vitro, or in vivo tissue samples, using a machine learning or other suitable learning algorithm.
Current imaging technologies struggle to detect certain tissue pathologies, such as non-enhancing brain tumors, due to the non-specific nature of existing imaging contrasts and the lack of voxel-wise histopathological validation. Studies have mainly focused on population-level correlations between images and biopsy samples, often overlooking individual tumor heterogeneity, which is crucial for personalized treatment and understanding tumor status. This invention addresses the need for systems and methods that can provide quantitative, non-invasive estimation of tissue histological features across different imaging contrasts.
The systems and methods enable the generation of histological feature maps from medical images, allowing for the quantification and correlation of individual histological and histopathological features with multiple imaging contrast mechanisms on a voxel-wise basis. Once trained, the model is not patient-specific and can be applied to any subject, facilitating non-invasive in vivo histology using only medical images and the trained model. The invention thus provides a framework for improved imaging and characterization of tissue pathology states in vivo, enhancing clinical interpretation and supporting the generation of comprehensive quantitative feature maps.
Claims Coverage
There are two independent claims in the patent, each describing a key inventive feature concerning the generation and training of histological feature maps from medical images using machine learning models.
Generating a histological feature map from medical images using a trained model
A method in which a computer system receives a plurality of medical images of a subject acquired with at least one medical imaging system. A trained model, developed using machine learning on training data comprising an image contrast matrix (from medical images) and a histological feature matrix (from quantitative histological feature values), is provided to the system. The system applies the trained model to the medical images to generate a histological feature map, which consists of quantitative values of a histological feature assigned to voxel locations within the subject.
Training a model with machine learning to predict quantitative histological feature values from medical images
A method involving providing a computer system with a plurality of medical images representing at least one medical image of each of several tissue samples, and corresponding quantitative histological feature values determined from these samples. Training data are formed as an image contrast matrix and a histological feature matrix. The computer system uses machine learning to train a model to predict quantitative values of histological features from medical images input to the model, and the trained model is stored.
The inventive features cover both the application of a pre-trained machine learning model to medical images for the generation of histological feature maps and the method for training such a model using carefully correlated imaging and histological data matrices.
Stated Advantages
Provides non-invasive, in vivo estimation of quantitative histological features using only medical images and a trained model.
Allows voxel-wise modeling and understanding of individual tissue pathology states, such as tumor heterogeneity, which is valuable for clinical decision-making.
Overcomes limitations of prior population-level analyses by enabling detailed, subject-specific characterization of tissue heterogeneity.
Supports improved interpretation of imaging biomarkers by explicitly relating imaging contrasts to underlying histological features.
Enables comprehensive, machine learning-based correlation of multiple imaging contrasts with histological features for enhanced tissue characterization.
The trained model, once developed, can be applied broadly to different subjects and is not patient-specific.
Documented Applications
Generation of histological feature maps, such as cell density maps, nuclei area maps, tumor grade maps (e.g., Gleason score), and quantitative maps of staining, from medical images of brain and prostate tissue.
Differentiation between sources of contrast enhancement in medical imaging, such as distinguishing pseudo-progression from viable tumor in oncology contexts.
Prediction and visualization of microscopic factors contributing to the appearance of brain tumors and other pathological tissue states using non-invasive imaging.
Clinical use for subject-specific, quantitative imaging of histopathological tissue characteristics, improving assessment and management of tumors or diseased tissues.
Application of the trained model to arbitrary subjects for extracting quantitative information about histological features from standard clinical imaging.
Interested in licensing this patent?