3D Rib Detection With Anatomical Queries for Adhesion Cases
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing rib segmentation methods in medical images, particularly CT images, suffer from poor accuracy due to issues like rib adhesion and structural damage, limiting their effectiveness in instance-level segmentation.
Innovation Solution
A neural network model, based on the transformer architecture, performs feature coding and parallel decoding to accurately segment ribs in medical images by learning the correspondence between rib query items and image regions, using a bipartite matching method to enhance rib detection and segmentation precision.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If existing rib segmentation methods are used, then the segmentation process can be completed, but the accuracy is poor due to rib adhesion and structural damage
Solution Approach 1:
The method segments the rib detection task into instance-level segmentation, dividing each rib into independent detectable units with unique identifiers. This allows the model to handle adhesion cases by treating each rib as a separate instance that can be individually localized and classified, overcoming the challenge of rib adhesion mentioned in the background.
Solution Approach 2:
The patent employs a detection box with 9 degrees of freedom that includes not only position and size parameters but also orientation parameters (rotation angles). This extended parameter set enables accurate representation of ribs with various orientations and structural variations, including damaged ribs, thereby improving segmentation accuracy despite structural damage and adhesion challenges.
2Measurement precision
If traditional segmentation methods are used, then the process is simpler, but the instance-level segmentation accuracy is poor
Solution Approach 1:
The patent introduces query items as intermediary elements that represent potential rib instances. These query items serve as mediators between the input medical image and the final segmentation output, allowing the model to systematically evaluate and identify each rib instance through attention mechanisms and feature matching, thereby achieving high instance-level accuracy.
Solution Approach 2:
The method extends the detection framework to handle instance-level segmentation by adding dimensional elements such as instance identifiers and orientation parameters. The detection box with 9 degrees of freedom adds rotational dimensions to the traditional bounding box concept, enabling precise instance-level segmentation while managing the increased complexity through structured parameterization.
3Measurement precision
If a detection box with 9 degrees of freedom is used, then rib position accuracy is improved, but the complexity of the detection system increases
Solution Approach 1:
The patent systematically parameterizes the detection box with 9 degrees of freedom, including 3 position parameters (x, y, z), 3 size parameters (width, height, depth), and 3 orientation parameters (rotation angles). This comprehensive parameterization enables precise localization and orientation of ribs in three-dimensional medical images, achieving high position accuracy while managing complexity through structured parameter definitions.
Data Source
AI summary
A method for detecting a rib with a medical image, a device and a medium are provided. Feature coding is performed, by using a neural network model, on different rib regions in the medical image to obtain multiple feature vectors. According to a learned corresponding relationship between rib query items and the different rib regions, a feature vector for respective rib query item is determined. Parallel decoding is performed on the feature vector for respective rib query item to obtain a rib detection result corresponding to respective rib query item. The rib query items are endowed with semantic information in anatomy, such that respective rib query items may focus on features of different rib regions, and thus realizing controllable rib detection in instance-level. Any rib detection result includes a rib prediction classification and a prediction value of a pose parameter, in a three-dimensional space, of a rib detection box.