Image Segmentation Training Using Union-Complement Label Loss
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image segmentation models face challenges due to inconsistent labeling results from multiple physicians, leading to uncertainty and suboptimal training outcomes using methods like union, intersection, or weighted average of labeling results, which result in unreasonable area boundaries and increased errors.
Innovation Solution
An image segmentation model training method that generates a complement area from the union of labeled target areas and uses it to calculate a first loss value, along with a second loss value from labeled target areas, to train the model, improving accuracy by learning to identify both target and non-target areas accurately.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If multiple physicians label the same medical image to provide diverse perspectives, then the comprehensiveness of labeling increases, but the inconsistency of labeling results increases
Solution Approach 1:
The patent segments the labeling process into two distinct components: the union area (areas labeled by at least one physician) and the complement area (areas not labeled by any physician). This segmentation allows the model to learn from both the agreement among physicians (union) and the uncertainty boundaries (complement), resolving the contradiction by treating consistency and comprehensiveness as separate learnable targets rather than conflicting objectives.
Solution Approach 2:
Instead of directly using the union area of labeled regions as training targets (conventional approach), the patent inverts the approach by using the complement area (areas outside the union) as a primary training target. This inversion allows the model to learn what should NOT be included in the segmentation, thereby improving boundary accuracy and reducing the impact of labeling uncertainty while maintaining comprehensiveness.
2Measurement precision
If the intersection of multiple labeling results is used for model training, then the consistency of training data increases, but the coverage of target areas decreases
Solution Approach 1:
The patent inverts the conventional approach of using intersection by utilizing the complement of the union instead. This inverted approach ensures that the model learns from the maximum possible coverage (union) while the complement provides clear negative examples, avoiding the coverage loss inherent in intersection-based methods.
3Area of stationary object
If the union of multiple labeling results is used for model training, then the coverage of target areas increases, but the labeling error increases
Solution Approach 1:
The patent segments the training objective into two parts: maximizing coverage through the union area and minimizing error through the complement area. By providing separate loss terms for these segmented objectives, the model can simultaneously achieve high coverage and low labeling error, resolving the contradiction between these two competing goals.
Solution Approach 2:
The complement area acts as an intermediary that mediates between coverage and error. It provides a clear, unambiguous training signal for areas that should definitely be excluded, serving as a mediator that guides the model to achieve high coverage while maintaining precision by clearly defining the boundaries of non-target areas.
4Measurement precision
If weighted average of labeling results is used to generate labeled image, then the gradient information is improved, but the influence of uncertainty boundary remains significant
Solution Approach 1:
The patent segments the training targets into discrete, non-overlapping regions (union area and complement area) rather than using continuous weighted averages. This segmentation eliminates the gradient information problem while also removing the influence of uncertainty boundaries, as the model learns from clear, binary distinctions rather than ambiguous weighted combinations.
Data Source
AI summary
The disclosure provides an image segmentation model training method and an electronic device. The method includes the following steps. A plurality of labeled images of a training image are obtained, wherein each of the labeled images includes a labeled target area and a background area. A union area of the labeled target areas of the plurality of labeled images is obtained, and a complement area of the union area is obtained. A predicted segmented image of the training image is generated using an image segmentation model. A first loss value is generated according to the predicted segmented image and the complement area. A second loss value is generated according to the predicted segmented image and the labeled target area of at least one of the plurality of labeled images. The image segmentation model is trained according to the first loss value and the second loss value.