Dense CNN for OCT Corneal Interface Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image analysis-based corneal interface segmentation approaches for Optical Coherence Tomography (OCT) images do not generalize across different OCT scanners, requiring manual parameter selection and being prone to inaccuracies due to speckle noise and low signal-to-noise ratios, leading to false positives and loss of discriminative features.
Innovation Solution
A deep-learning model with dense connections between layers of a convolutional neural network (CNN) is established, incorporating downsampling and upsampling techniques using max pooling and nearest neighbor interpolation, trained on images from multiple OCT scanners to segment corneal interfaces like Epithelium, Bowman's Layer, and Endothelium accurately.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If existing image analysis-based corneal interface segmentation approaches are used, then manual parameter selection is required, but the approaches do not generalize to volumes acquired from different OCT scanners and are prone to inaccuracies
Solution Approach 1:
The patent transforms the segmentation approach from manual parameter selection to automated deep learning-based parameter extraction. The CNN model learns optimal segmentation parameters automatically from training data across multiple OCT scanner types, enabling the system to adapt to different scanners without manual reconfiguration while maintaining high segmentation accuracy through learned feature representations
Solution Approach 2:
The patent creates a universal segmentation model that works across different OCT scanner types by training on diverse training datasets from multiple scanners. The deep learning model achieves multi-functionality by generalizing to volumes from different scanners without requiring scanner-specific manual parameter tuning, thereby improving both adaptability and precision simultaneously
2Ease of operation
If manual parameter selection is used for segmentation, then the process is ad hoc, but it leads to false positives and loss of discriminative features
Solution Approach 1:
The patent replaces manual parameter selection (mechanical/human operation) with an automated deep learning system. The CNN model automatically extracts segmentation parameters from input images, eliminating the need for manual intervention and reducing false positives through consistent, data-driven parameter optimization that preserves discriminative features
Solution Approach 2:
The segmentation system performs self-service by automatically selecting optimal parameters through the trained CNN model without requiring manual intervention. The model learns from training data and autonomously applies segmentation parameters to new images, improving ease of operation while maintaining reliability through consistent automated decision-making
3Measurement precision
If dense connections are established between layers of CNN, then faint boundary features are maintained and retrieval is improved, but the device complexity increases
Solution Approach 1:
The patent implements dense connections where each layer is nested with connections to all preceding layers, creating a nested architecture that allows feature maps from earlier layers to be accessible at deeper layers. This nested structure enables the network to maintain and retrieve faint boundary features by combining low-level edge detections with high-level semantic information, improving detection precision despite increased architectural complexity
Data Source
AI summary
Provided is a system, method, and computer program product for creating a deep-learning model for processing image data. The method includes establishing dense connections between each layer of a plurality of layers of a convolutional neural network (CNN) and a plurality of preceding layers of the CNN, downsampling an input of each downsampling layer of a plurality of downsampling layers in a first branch of the CNN, and upsampling an input of each upsampling layer of a plurality of upsampling layers in a second branch of the CNN by convolving the input.