OCT Biomarker Detection Using 3D Slice Tiling and Transfer Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for detecting OCT risk factors for progression to late AMD in clinical settings are unreliable due to the reliance on manual inspection by clinicians, which is laborious and prone to human error, and existing deep learning approaches require large datasets that are not readily available for widespread clinical application.
Innovation Solution
A deep learning method that reshapes 3D OCT images into 2D slices, applies a pre-trained feature extractor, and uses a convolutional neural network to automatically detect biomarkers, leveraging transfer learning from external datasets like ImageNet to overcome data scarcity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If manual inspection by clinicians is used to detect OCT risk factors, then clinical judgment and adaptability are maintained, but labor intensity increases and reliability decreases
Solution Approach 1:
The patent replaces manual clinical inspection with an automated deep learning system that uses convolutional neural networks to detect AMD risk factors from OCT images. The system automatically processes images through multiple CNN layers to identify biomarkers such as drusen, pigment epithelial detachments, and subretinal neovascularization, eliminating the labor-intensive manual review process while maintaining detection reliability.
Solution Approach 2:
The system performs self-service by automatically detecting and classifying AMD risk factors without requiring clinician intervention for each specific detection task. The deep learning model independently analyzes OCT images, extracts features, and generates diagnostic predictions, allowing the system to serve itself in the detection process rather than requiring continuous human guidance for each imaging analysis.
2Productivity
If deep learning approaches are used to automatically detect biomarkers, then productivity and detection speed improve, but data requirements increase
Solution Approach 1:
The patent applies preliminary action by pre-training the convolutional neural network on large datasets of natural images (such as ImageNet) before training on the specific medical OCT imaging data. This pre-training phase establishes foundational feature extractors that can recognize basic visual patterns, allowing the model to achieve high detection accuracy for AMD biomarkers without requiring equally large datasets specific to ocular imaging.
Solution Approach 2:
The system achieves universality by using a pre-trained CNN architecture that was originally designed for general image classification tasks and adapting it to specialize in medical imaging. The same pre-trained network can detect multiple different AMD biomarkers and apply to various OCT imaging scenarios, making the system versatile and reducing the need for separate specialized models for each detection task.
3Loss of information
If 3D OCT images are processed directly, then spatial information is preserved, but computational complexity increases
Solution Approach 1:
The patent applies segmentation by dividing the 3D OCT volume into multiple 2D axial slices that are processed independently through the convolutional neural network. Each slice is a separate 2D image that can be analyzed by the pre-trained CNN, allowing the system to maintain spatial information across the retinal layers while avoiding the computational burden of processing the entire 3D volume as a single complex data structure.
Solution Approach 2:
The system transforms the 3D OCT data from a volumetric representation into a series of 2D axial slices, changing the dimensional representation to match the input format expected by pre-trained 2D convolutional neural networks. This dimensional transformation allows efficient processing using established 2D CNN architectures while preserving the spatial relationships and structural information present in the original 3D OCT images.
Data Source
AI summary
Deep learning methods and systems for detecting biomarkers within optical coherence tomography volumes using such deep learning methods and systems are provided. Embodiments predict the presence or absence of clinically useful biomarkers in OCT images using deep neural networks. The lack of available training data for canonical deep learning approaches is overcome in embodiments by leveraging a large external dataset consisting of foveal scans using transfer learning. Embodiments represent the three-dimensional OCT volume by “tiling” each slice into a single two dimensional image, and adding an additional component to encourage the network to consider local spatial structure. Methods and systems, according to embodiments are able to identify the presence or absence of AMD-related biomarkers on par with clinicians. Beyond identifying biomarkers, additional models could be trained, according to embodiments, to predict the progression of these biomarkers over time.