Compensating OCT Image Segmentation Lines Using CNN Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current image segmentation methods for OCT images, such as linear interpolation, often result in severed boundaries due to similar or dissimilar neighboring pixel colors, leading to inaccurate compensation of segmentation lines, which can hinder precise diagnosis and treatment in medical imaging.
Innovation Solution
A method using a classification model with a convolutional neural network (CNN) architecture and a predetermined kernel to predict the y-coordinate trend of next pixels, connecting each pixel in the x and y directions, thereby compensating for missing pixels and learning features and positions for more accurate segmentation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If linear interpolation is used to compensate for severed boundaries, then the compensation process is simple, but the segmentation line continuity is poor when pixel colors are similar or too different
Solution Approach 1:
The patent changes the compensation method from linear interpolation to a classification model-based approach. The classification model predicts the y-coordinate trend of next pixels by learning from training data, dynamically adjusting the compensation strategy based on pixel color characteristics. This resolves the contradiction by maintaining simplicity through automated classification while achieving continuity through learned patterns that adapt to similar or different pixel colors.
Solution Approach 2:
The patent replaces the mechanical linear interpolation process with a neural network-based classification system. Instead of using fixed mathematical interpolation formulas, the system uses a trained classification model to predict pixel trends, substituting rigid mechanical computation with adaptive intelligent processing that maintains both simplicity and precision.
2Ease of operation
If conventional segmentation models are used, then the segmentation process is straightforward, but severed boundaries occur in the segmentation results
Solution Approach 1:
The patent introduces a classification model as an intermediary between the conventional segmentation model and the final segmentation result. This intermediary component predicts the y-coordinate trend of next pixels and compensates for severed boundaries, maintaining the straightforwardness of the original segmentation process while improving reliability through the additional prediction step.
Solution Approach 2:
The patent performs preliminary classification and prediction of pixel trends before finalizing the segmentation result. By using the classification model to predict next pixel positions in advance, the system prevents severed boundaries from occurring in the final output, maintaining both operational simplicity and result reliability.
3Measurement precision
If a classification model with convolution operation is used to predict next pixel positions, then segmentation accuracy is improved, but the device complexity increases
Solution Approach 1:
The patent segments the complex task into two distinct parts: a classification model for predicting y-coordinate trends and a convolution operation for processing segmentation images. This segmentation of functions allows each component to be optimized independently, improving overall segmentation accuracy while managing complexity through modular design.
Solution Approach 2:
The patent adds a new dimension to the segmentation process by introducing a classification model that operates in the pixel prediction domain. Instead of only processing spatial dimensions, the system incorporates a temporal/predictive dimension by forecasting next pixel positions, thereby improving accuracy without proportionally increasing complexity.
Data Source
Figure 1~2
Figure 3
Figure 4
AI summary
A method and an apparatus for compensating image segmentation lines (20) are provided. In the method, image segmentation is applied to a medical image (42) captured by using a segmentation model (44) to obtain a segmentation image (46) including at least one segmentation line (12a) between multiple layers in the medical image. Convolution computation is then performed on the segmentation image (46) by using a kernel (60, 60a) of a trained classification model (50) to predict a location of a next pixel connected to a current pixel in the respective segmentation line (12a) within the segmentation image (46), in which the pixel to be predicted is limited to a neighboring pixel of the current pixel in a prediction direction. The predicted pixels are connected to form a compensated segmentation line (14a) for each segmentation line (12a).