Multi-Sieving CNN for Microaneurysm Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for detecting microaneurysms in retinal images face challenges due to variations in image lighting, clarity, and noise interference, with shallow neural networks and support vector machines lacking sensitivity and stability, and existing methods relying on priori knowledge and assumptions.
Innovation Solution
A method using a multi-sieving convolutional neural network (CNN) that partitions images with random ferns to create auxiliary channel images, combining text diagnosis results with image information for training, and iteratively refining the model to improve classification accuracy and stability.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If shallow neural networks or support vector machines are used for microaneurysm detection, then the detection method is simpler and faster, but the sensitivity and stability are insufficient
Solution Approach 1:
The patent implements a multi-stage detection framework where coarse detection and fine detection modules are nested sequentially. The coarse detection module first identifies potential microaneurysm regions, which are then fed into the fine detection module for precise classification. This nested structure enables the system to achieve high sensitivity and stability comparable to complex models while maintaining reasonable computational efficiency.
Solution Approach 2:
The detection task is segmented into two distinct stages: coarse detection for initial screening and fine detection for precise classification. Each stage uses optimized network architectures and loss functions tailored to its specific objectives, allowing the system to achieve high overall performance without requiring a single monolithic complex model.
2Measurement precision
If filter-based or morphological-based methods are used, then the detection process is faster, but the accuracy deteriorates due to interference from other red lesions and noise
Solution Approach 1:
The coarse detection module performs preliminary filtering of potential microaneurysm regions before the fine detection module processes them. This preliminary action reduces the search space and eliminates obvious false positives early in the pipeline, allowing the system to achieve high accuracy without requiring exhaustive processing of the entire image.
Solution Approach 2:
The fine detection module applies specialized processing only to the regions identified by coarse detection, using localized feature extraction and classification. This local quality approach concentrates computational resources on critical regions, achieving high detection accuracy while minimizing overall computation time.
3Adaptability or versatility
If methods based on priori knowledge and strong assumptions are used, then the model training is simpler, but the adaptability to different image conditions deteriorates
Solution Approach 1:
The patent employs adaptive loss functions and data augmentation strategies that automatically adjust to different image conditions during training. The loss function weights and augmentation parameters are dynamically modified based on the specific characteristics of the training data, enabling the model to adapt to varying lighting, noise levels, and image qualities without requiring manual reconfiguration.
Solution Approach 2:
The model performs self-supervised learning through automated annotation and iterative refinement. The system automatically generates training labels from detected results and continuously improves its performance through self-correction, reducing the need for manual intervention and expert knowledge while enhancing adaptability to different clinical scenarios.
Data Source
AI summary
A method of automatically detecting microaneurysm based on multi-sieving convolutional neural network (CNN), includes the following steps of: A1), partitioning an image to be detected using random fern and obtaining an auxiliary channel image of the image according to a first partition result; and A2), inputting the auxiliary channel image obtained from step A1) and the image to a multi-sieving CNN training model to perform a detection and obtaining a microaneurysm detection result of the image. The process of establishing the training model includes: B1), using a current microaneurysm diagnostic report as samples and partitioning a lesion image in the microaneurysm diagnostic report using the random fern, and establishing the auxiliary channel image according to a second partition result; B2), comparing the obtained auxiliary channel image with a lesion-marked image of pixels, clarifying the samples according to a comparing result and performing the multi-sieving CNN training.