Contrastive Gated Attention for Small Object Detection in OCT Images
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current AI object detection methods face challenges in accurately identifying and localizing small, highly heterogeneous objects in medical imaging, such as OCT scans, due to limited training datasets and noisy signals from misaligned structured information, leading to poor performance in detecting small objects.
Innovation Solution
A computer-implemented method using contrastive representation learning with a contrastive-induced gated attention (CIGA) module, which processes pairs of positive and negative images to extract feature vectors, contrasts them, and gates the results to form an attention vector, enhancing the detection of small objects by focusing on important features and reducing noise from misalignment.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If deep learning methods are used for object detection in OCT images, then automation and efficiency are improved, but performance is poor due to small datasets and heterogeneous features
Solution Approach 1:
The patent introduces an intermediary alignment module that processes structured information (anatomical layers, vascular structures) before feeding it to the detection network. This intermediary component aligns the structured data with image features, mediating between the raw input and the detection algorithm, thereby improving reliability without sacrificing automation.
Solution Approach 2:
The patent performs preliminary alignment and integration of structured information before the main detection process. By pre-processing the structured anatomical data to match image coordinates and semantics, the system prepares optimal input conditions for the detection network, enabling better performance with limited training data.
2Adaptability or versatility
If standard object detection algorithms are applied to OCT images, then general applicability is maintained, but detection precision for small objects deteriorates
Solution Approach 1:
The patent applies local quality enhancement by integrating structured information that provides local anatomical context to specific regions of the image. The structured data (layer positions, vascular patterns) adds localized semantic information that helps the detector distinguish small objects in their specific anatomical context, improving precision without requiring a completely different algorithm.
Solution Approach 2:
The patent creates a composite representation by combining standard image features with structured anatomical information. This composite input integrates multiple data modalities (pixel intensities, layer structures, vascular patterns) into a unified feature space, enabling the standard detection algorithm to achieve higher precision for small objects.
3Quantity of substance
If deep learning networks are trained on limited medical imaging datasets, then data availability constraints are respected, but learning robust features becomes difficult
Solution Approach 1:
The structured information acts as an intermediary that provides additional supervisory signals to the learning process. By incorporating aligned anatomical structures as auxiliary training targets, the network learns more robust features from fewer images, as the structured data provides consistent semantic guidance across the limited dataset.
Solution Approach 2:
The structured information serves multiple functions simultaneously: it provides anatomical context for detection, serves as an additional training signal, and offers semantic alignment guidance. This multi-functionality allows the system to extract maximum learning value from limited training data, improving feature robustness without requiring more images.
4Ease of manufacture
If bounding box labeling is used for object detection, then labeling complexity is reduced compared to segmentation, but localization precision may be insufficient for small objects
Solution Approach 1:
The patent enhances bounding box detection by adding local quality information from structured data. The aligned anatomical structures provide local context that helps the network more accurately localize small objects within their bounding boxes, improving precision without requiring more complex segmentation-level labeling.
Data Source
AI summary
A computer-implemented method of enhancing object detection in a digital image of known underlying structure using pre-processed images with underlying structure and with any objects detected and bounding boxes inserted over the objects, the method comprising: extracting or generating images with the underlying structure but without objects detected as negative images; extracting images with the underlying structure and with an object detected as positive images; inputting pairs of negative and positive images through a feature extraction section in a neural network to extract feature vectors of the images; contrasting feature vectors of each pair of positive and negative images and thereby provide a contrast vector and gating the result to form an attention vector; processing the attention vector and the feature vector of the positive image to produce an output.