Semantic Segmentation for Minute Point and Line Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image processing systems using convolutional neural networks (CNNs) struggle to accurately identify minute objects such as the center point of a vertebral body, center points of eyeballs, or the center line of the aorta in medical images.
Innovation Solution
A semantic segmentation model is trained with a greater weight assigned to first pixels corresponding to objects, using annotation images where these pixels are set to a different value than other pixels, and calculates a loss to identify discrete points or lines in medical images, employing probability distribution maps and non-maximum suppression processing to enhance accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a standard semantic segmentation model is used to identify objects in medical images, then large objects such as lung fields and spines can be identified with relatively high accuracy, but minute objects such as vertebral body center points, eyeball center points, and aorta center lines cannot be accurately identified
Solution Approach 1:
The patent applies local quality by assigning different weights to different pixels in the loss function. Specifically, pixels corresponding to minute objects (points or lines) are assigned higher weights than other pixels, enabling the model to focus computational attention on critical small-scale features while maintaining overall segmentation performance. This resolves the contradiction by enhancing local identification accuracy without requiring a complete redesign of the overall model architecture.
Solution Approach 2:
The patent modifies the loss function parameters by introducing position-dependent weighting factors that emphasize pixels corresponding to minute objects. This parameter change allows the model to prioritize learning features relevant to small-scale structures during training, thereby improving identification accuracy for minute objects without significantly increasing model complexity.
2Measurement precision
If the model focuses on identifying minute objects by assigning greater weight to corresponding pixels, then identification accuracy of minute objects improves, but training time and computational resources increase
Solution Approach 1:
The patent applies partial action by selectively weighting only those pixels that correspond to minute objects, rather than uniformly increasing weight across all pixels. This targeted approach concentrates computational effort where most needed, improving minute object identification accuracy while minimizing the overall increase in training time and computational resources.
3Measurement precision
If annotation images are created with distinct pixel values for objects versus non-objects, then training effectiveness improves, but annotation complexity and time requirements increase
Solution Approach 1:
The patent applies segmentation by dividing the annotation process into distinct categories: pixels corresponding to points are assigned one value, pixels corresponding to lines are assigned another value, and other pixels are assigned a third value. This segmented annotation approach simplifies the overall process by providing clear, distinct categories that are easier to annotate consistently while maintaining training effectiveness.
Data Source
AI summary
A processor uses a semantic segmentation model that has been trained using an annotation image in which a first pixel corresponding to at least any one of one point corresponding to an object, a plurality of discrete points corresponding to a plurality of objects, or a line corresponding to an object having a line structure is set as a first pixel value and a second pixel other than the first pixel is set as a second pixel value different from the first pixel value, the model having been trained by assigning a greater weight to the first pixel than to the second pixel to calculate a loss, inputs an image to the model and outputs a feature amount map having a feature amount related to the one point, etc. in the image from the model, and identifies the one point, etc. in the image based on the map.