Retinal Anomaly Detection via Multi-Resolution CNN Ensemble
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image analysis systems for eye diseases face challenges in accurately identifying multiple diseases co-existing in one eye due to insufficient training data, especially for rare diseases, and rely on normative data which can be biased, leading to unreliable responses for non-target diseases.
Innovation Solution
A multivariate and multi-resolution retinal image anomaly detection system using convolutional neural networks (CNNs) and support vector machines (SVMs) that processes B-scan images from 3D OCT volumes, incorporating pre-processing techniques like image flattening and noise removal, and employs ensemble anomaly detection combining classification and inpainting methods to identify structural changes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If normative data is used as training data for anomaly detection, then the system can screen abnormal eyes effectively, but it gives unreliable responses to non-target class disease data and cannot accurately identify multiple co-existing diseases
Solution Approach 1:
The system segments the anomaly detection task into multiple specialized binary classifiers, each trained to detect a specific disease type. Instead of using a single normative-based detector, the system divides the problem into separate detection modules for different disease categories, allowing each classifier to specialize in one disease type while maintaining overall system reliability and versatility.
Solution Approach 2:
The system implements a universal ensemble framework that integrates multiple disease-specific classifiers into a single multi-functional system. This ensemble approach allows the system to handle multiple disease types simultaneously, providing both reliable anomaly detection and versatile disease identification capabilities through the collective power of specialized classifiers.
2Adaptability or versatility
If multiple binary classifiers are used to detect different diseases, then the system can identify specific disease types, but it requires significantly large amounts of disease training data which is difficult to obtain for rare diseases
Solution Approach 1:
The system performs preliminary action by pre-training deep convolutional neural networks on large-scale normative data before deploying them as disease-specific classifiers. This pre-training establishes strong feature extraction capabilities that can be adapted to detect various disease types with limited pathological training data, reducing the data requirement burden for rare diseases.
Solution Approach 2:
The system changes parameters by adjusting the training configuration and data requirements for each disease-specific classifier based on data availability. For rare diseases with limited training data, the system modifies training parameters and leverages transfer learning from normative data, allowing effective classifier development without requiring large amounts of rare disease training samples.
3Device complexity
If a single binary classifier is used to distinguish one major disease from normalcy, then the system is simple to implement, but it gives unreliable responses to non-target class disease data and cannot handle multiple co-existing diseases
Solution Approach 1:
The system segments the single classifier into multiple specialized binary classifiers, each responsible for detecting a specific disease type. This segmentation improves reliability for non-target diseases by ensuring each classifier focuses on its designated disease, while the modular structure maintains implementation simplicity through clear division of responsibilities.
Solution Approach 2:
The system merges multiple specialized binary classifiers into an integrated ensemble system that maintains simplicity in implementation while improving reliability. The ensemble combines the strengths of individual classifiers through weighted voting or averaging, providing robust responses for both target and non-target diseases while preserving the simplicity of binary classification approaches.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Machine learning technologies are used to identify and separating abnormal and normal subjects and identifying possible disease types with images (e.g., optical coherence tomography (OCT) images of the eye), where the machine learning technologies are trained with only normative data. In one example, a feature or a physiological structure of an image is extracted, and the image is classified based on the extracted feature. In another example, a region of the image is masked and then reconstructed, and a similarity is determined between the reconstructed region and the original region of the image. A label (indicating an abnormality) and a score (indicating a severity) can be determined based on the classification and/or the similarity.