Diagnosis of a disease condition using an automated diagnostic model
Inventors
Abramoff, Michael • Quellec, Gwenole
Assignees
University of Iowa Research Foundation UIRF • US Department of Veterans Affairs
Publication Number
US-11935235-B2
Publication Date
2024-03-19
Expiration Date
2031-12-06
Interested in licensing this patent?
MTEC can help explore whether this patent might be available for licensing for your application.
Abstract
A method of identifying an object of interest can comprise obtaining first samples of an intensity distribution of one or more object of interest, obtaining second samples of an intensity distribution of confounder objects, transforming the first and second samples into an appropriate first space, performing dimension reduction on the transformed first and second samples, whereby the dimension reduction of the transformed first and second samples generates an object detector, transforming one or more of the digital images into the first space, performing dimension reduction on the transformed digital images, whereby the dimension reduction of the transformed digital images generates one or more reduced images, classifying one or more pixels of the one or more reduced images based on a comparison with the object detector, and identifying one or more objects of interest from the classified pixels.
Core Innovation
The invention provides methods and systems for optimal, user-friendly object and background separation in digital images, particularly aimed at detecting characteristics such as lesions evidencing disease. It introduces a new, fast approach to detect objects in images by developing optimized detectors through a systematic framework that can be applied in various domains including retinal imaging and other medical or computer vision applications.
The problem addressed is that prior methods for automated detection of diseases such as diabetic retinopathy and age-related macular degeneration inadequately differentiate lesions from confounding structures like retinal blood vessels or lesions proximate to vessels, resulting in false positives or missed detections. Existing detectors are not optimal, and current detection performance requires enhancements before clinical application.
The invention solves this by providing an optimal filter framework that automatically generates features to distinguish target lesions, including atypical ones, from negative and positive lesion confounders. This framework employs expert-driven mathematical modeling or data-driven annotated image sampling to create representative samples, performs dimension reduction to generate an object detector, and classifies image pixels accordingly, enabling fast and accurate lesion detection and differentiation. This systematic approach closely mimics visual processing and achieves near instantaneous detection useful for screening and diagnostic feedback.
Claims Coverage
The patent includes multiple independent claims covering a system and method for diagnosing disease conditions using trained object detection and diagnostic models applied to images.
System for diagnosis using trained object detection and automated diagnostic models
The system comprises one or more processors and computer readable media storing an object detection model component trained to output probabilities of objects of interest in input images, an automated diagnostic model component trained to output probabilistic disease diagnoses based on those probabilities, and instructions to process input images through these components to output a diagnosis.
Applying supervised procedure with trained filters to image samples
The object detection employs a supervised procedure comprising a plurality of trained filters configured to identify the probability of objects of interest indicative of disease within obtained image samples from different regions.
Training the object detection and diagnostic models via annotated examples
The supervised procedure and diagnostic model are trained by fitting to plural training examples consisting of images annotated with object locations, or images with object presence probabilities and disease labels.
Method of diagnosing disease by cascading object detection and diagnostic models
The method receives an input image, processes it through a trained object detection model producing object presence probabilities at image locations, passes these probabilities to a trained automated diagnostic model, and outputs a probabilistic diagnosis of disease based on this processing.
The independent claims collectively describe a system and method that utilize supervised learning to generate trained filters detecting disease-related objects in images, then use these detections to provide probabilistic disease diagnoses, with training based on annotated input images. This approach captures the inventive combination of object detection and diagnostic modeling components integrated for automated disease diagnosis from images.
Stated Advantages
The method enables reproducible, efficient, and early detection or screening of disease conditions in large at-risk populations.
The optimal filter framework achieves detection performance comparable to or better than previous algorithms, while being substantially faster, allowing processing of retinal images in less than a second.
Instantaneous feedback is possible to operators and patients, enabling immediate diagnostic assessment.
The approach requires limited expert knowledge and minimal annotation effort, supporting adaptability to new datasets or domains.
The system closely mimics visual information processing by the brain, improving lesion detection and differentiation from confounders.
Documented Applications
Detection and diagnosis of retinal diseases such as diabetic retinopathy and age-related macular degeneration through identification of lesions like microaneurysms and drusen.
Differentiation between similar appearing lesions, for example, drusen versus flecks related to Stargardt's disease.
Retinal vessel bifurcation detection for clinical evaluation and downstream image processing.
General object detection in diverse images such as blood cells, fruit trees, satellite images of helicopter landing pads, and internet images of objects like apples.
Interested in licensing this patent?