Lesion Clustering for Pathology Prediction in Medical Images
Solution Overview
Problem
Automated classifiers require fixed data arrays to predict disease states or pathological conditions, which is problematic when dealing with variable numbers of lesions in medical images, leading to the omission of spatial distribution information.
Innovation Solution
A computer-implemented method and apparatus that acquires lesion locations, applies clustering algorithms to identify lesion clusters, and categorizes them into predetermined categories, providing a fixed number of data outputs for classification, allowing for the determination of pathology levels in diabetic retinopathy and age-related macular degeneration without requiring fixed-sized data arrays.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If automated classifiers use fixed data arrays to predict disease states, then the classification process is simplified and standardized, but spatial distribution information of lesions is omitted when dealing with variable numbers of lesions
Solution Approach 1:
The patent segments the retinal image into multiple predefined regions of interest (such as macula, perimacular, and peripheral zones). Lesions are then counted and analyzed within each segment separately, transforming the variable spatial distribution problem into a structured multi-region analysis that preserves location information while providing fixed regional counts for classification.
Solution Approach 2:
The patent adds a spatial dimension by mapping lesion locations to specific retinal regions before aggregation. Instead of simply counting total lesions, the system analyzes the distribution pattern across different spatial zones, converting 2D spatial information into a structured multi-dimensional feature set that maintains both quantity and location data.
2Measurement precision
If the system processes variable numbers of lesion data points, then accuracy in representing actual pathology is improved, but the data structure becomes complex and difficult to process
Solution Approach 1:
The patent divides the complex variable lesion data into discrete regional categories (e.g., macular lesions, perimacular lesions, peripheral lesions). This segmentation transforms the unstructured variable list of lesions into a structured set of regional counts, simplifying the data structure while preserving accuracy through location-specific analysis.
Solution Approach 2:
The patent applies different analysis criteria to different retinal regions based on their clinical significance. Each region is evaluated with appropriate local parameters (such as lesion density in the macula versus peripheral distribution), allowing accurate pathology representation while maintaining a unified processing framework through region-specific customization.
Data Source
Figure 1~2
Figure 3
Figure 4~10
AI summary
A computer-implemented method, system, and computer-readable medium, for determining a pathologic condition from a medical image of a portion of a subject. The method includes acquiring a plurality of lesion locations in the medical image; applying a clustering algorithm to the plurality of lesion locations to identify at least one lesion cluster and corresponding lesion cluster data; categorizing each lesion cluster into one of a set of predetermined categories based on the identified lesion cluster data; applying at least one function to the lesion cluster data with regard to each category of the set of predetermined categories, wherein the at least one function provides a fixed number of data outputs; and determining a pathologic condition by processing the fixed number of data outputs of each category of the set based on a classification algorithm trained on image data defining medical images of the portion of a plurality of subjects.