Lesion Segmentation via 2D-3D Feature Fusion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The manual detection and delineation of brain metastases in MRI images is time-consuming, expensive, and prone to intra- and inter-expert variability, while existing deep learning methods lack sufficient sensitivity for clinical applications.
Innovation Solution
A system that combines a 2D segmentation network and a 3D segmentation network to automatically segment lesions from medical images, fusing 2D and 3D features to enhance detection sensitivity and specificity, using a trained machine learning-based fusion model.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual detection and delineation of brain metastases is performed by neurologists or radiologists, then detection accuracy can be maintained at expert level, but the process is time-consuming and expensive
Solution Approach 1:
The patent applies deep learning models (2D and 3D segmentation networks) that learn from and replicate the expertise of radiologists to automatically detect and segment brain metastases. The system processes MRI images through neural networks that have been trained on annotated data, producing segmentation masks that replicate expert detection accuracy while eliminating the time consumption associated with manual analysis.
2Measurement precision
If manual detection and delineation is performed by neurologists or radiologists, then detection accuracy can be maintained at expert level, but inter-expert variability and intra-expert variability occur
Solution Approach 1:
The patent transforms the detection process from human-based to algorithm-based by changing the fundamental parameter of the detector. The deep learning models process images through fixed mathematical operations and learned features, eliminating the variability inherent in human interpretation. The system applies consistent segmentation criteria across all images, ensuring reproducible results that do not depend on which radiologist performs the analysis.
3Extent of automation
If conventional deep learning methods are used for automatic detection, then automation is achieved, but detection sensitivity is too limited for clinical applications
Solution Approach 1:
The patent divides the detection task into separate 2D and 3D processing streams. The 2D segmentation network analyzes individual slices for high-resolution local features, while the 3D segmentation network processes volumetric data for spatial context and through-plane relationships. These segmented results are then fused to produce the final detection, allowing each component to specialize in specific aspects of lesion detection, thereby achieving high sensitivity suitable for clinical use.
Solution Approach 2:
The patent transitions from conventional 2D image analysis to integrated 3D volumetric analysis by adding the depth dimension to the detection process. The 3D segmentation network processes stacked 2D slices as a volumetric dataset, enabling detection of lesions that span multiple slices and providing contextual information about three-dimensional spatial relationships. This dimensional expansion significantly improves detection sensitivity compared to traditional 2D-only approaches.
4Measurement precision
If 2D and 3D segmentation networks are combined, then detection sensitivity and specificity are improved, but system complexity increases
Solution Approach 1:
The patent segments the complex detection system into two independent but complementary parts: a 2D segmentation network for local detail analysis and a 3D segmentation network for spatial context analysis. Each network can be trained, optimized, and processed separately before their results are fused. This modular segmentation of the system allows for manageable complexity in each component while achieving superior overall performance through their coordinated operation.
Data Source
AI summary
Systems and methods for automatic segmentation of lesions from a 3D input medical image are provided. A 3D input medical image depicting one or more lesions is received. The one or more lesions are segmented from one or more 2D slices extracted from the 3D input medical image using a trained 2D segmentation network. 2D features are extracted from results of the segmentation of the one or more lesions from the one or more 2D slices. The one or more lesions are segmented from a 3D patch extracted from the 3D input medical image using a trained 3D segmentation network. 3D features are extracted from results of the segmentation of the one or more lesions from the 3D patch. The extracted 2D features and the extracted 3D features are fused to generate final segmentation results. The final segmentation results are output.