Fundus Image Disc Hemorrhage Detection Using Dual Machine Learning Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for detecting disc hemorrhage in fundus images face challenges in accuracy and reproducibility due to variations in pixel values and the difficulty in distinguishing hemorrhagic areas within the optic disc region, leading to inconsistent diagnosis.
Innovation Solution
An ophthalmic information processing apparatus utilizing machine learning models for disc hemorrhage determination and region detection, combining a disc hemorrhage determination model and a disc hemorrhage region detection model to accurately identify and analyze disc hemorrhage regions in fundus images, with optional integration of OCT data for precise positioning and shape analysis.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If conventional methods using pixel value differences are used to extract hemorrhage regions, then the processing is simple, but the measurement precision and reliability of disc hemorrhage detection deteriorates due to variations in pixel values and difficulty in distinguishing hemorrhagic areas
Solution Approach 1:
The patent replaces conventional pixel value difference methods with machine learning models (CNN and U-Net) that automatically learn optimal feature representations from labeled fundus images. The deep learning models process the entire image structure rather than relying on simple pixel arithmetic, achieving superior detection precision while maintaining computational efficiency through automated feature extraction.
Solution Approach 2:
The patent transitions from fixed pixel value thresholds to adaptive machine learning models that automatically adjust detection parameters based on training data. The models learn optimal feature weights and activation thresholds during training, enabling them to handle pixel value variations and lighting conditions without manual parameter adjustment, thus improving measurement precision.
2Measurement precision
If doctors manually observe fundus images to determine disc hemorrhage, then the diagnostic accuracy can be high with experience, but the productivity and consistency deteriorate due to subjectivity and variability in human diagnosis
Solution Approach 1:
The patent implements self-service through automated machine learning models that independently perform disc hemorrhage detection without requiring doctor intervention. The deep learning systems automatically analyze fundus images, making diagnoses consistent and reproducible while maintaining high accuracy. The models learn from training data and apply the same decision logic to all cases, eliminating human variability.
Solution Approach 2:
The patent incorporates feedback mechanisms where the model performance is continuously evaluated and improved through training on labeled images. The system learns from correct and incorrect detections, adjusting its internal parameters to improve accuracy over time. This feedback loop ensures consistent high-quality diagnoses while maintaining productivity.
3Device complexity
If a single machine learning model is used for both determination and region detection, then the device complexity is reduced, but the measurement precision and reliability of hemorrhage region analysis deteriorates
Solution Approach 1:
The patent segments the detection task into two specialized models: a classification model (CNN) for determining presence/absence of disc hemorrhage and a segmentation model (U-Net) for precisely identifying hemorrhage regions. This division of labor allows each model to be optimized for its specific function, with the segmentation model focusing entirely on region boundaries and characteristics, thereby improving measurement precision while the modular architecture keeps overall system complexity manageable.
Solution Approach 2:
The patent adds a dimensional separation between the determination task (binary classification) and region detection task (pixel-level segmentation). By operating in different functional dimensions with specialized models, the system achieves higher precision in region analysis without excessive complexity. The segmentation model operates at a finer granularity level, providing detailed regional information that complements the coarser determination output.
Data Source
AI summary
An ophthalmic information processing apparatus includes a determiner and a detector. The determiner is configured to determine a presence or absence of a disc hemorrhage for a front image of a fundus of a subject's eye, using a disc hemorrhage determination model obtained by performing machine learning using a plurality of fundus images labeled with labels indicating the presence or absence of a disc hemorrhage as first teaching data. The detector is configured to detect a disc hemorrhage region depicted in the front image that is determined to have the disc hemorrhage, using a disc hemorrhage region detection model obtained by performing machine learning using a plurality of pairs of image groups as second teaching data, each pair having a front image of a fundus and a disc hemorrhage region image representing a disc hemorrhage region depicted in the front image.