Marginal Space Deep Neural Networks for Anatomical Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing anatomical object detection methods in medical images, such as Marginal Space Learning (MSL), are not robust for detecting anatomical objects with large variations in anatomy, shape, or appearance, as they rely on handcrafted image features that are inefficient in capturing complex distributions and are computationally complex for large 2D+time or 3D volumetric images.
Innovation Solution
The use of marginal space deep neural networks that divide the parameter space into series of marginal search spaces with increasing dimensionality, where deep neural networks are trained directly on image data to learn high-level domain-specific features for anatomical object detection, rather than relying on handcrafted features, allowing for discriminative or regression-based detection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If handcrafted image features are used for anatomical object detection, then the detection process is simpler to implement, but the detection robustness deteriorates for anatomical objects with large variations in anatomy, shape, or appearance
Solution Approach 1:
The patent replaces handcrafted feature extraction methods with deep neural networks that automatically learn image features. The deep neural network learns hierarchical features directly from raw image data, substituting the manual feature engineering process with an automated learning system that adapts to variations in anatomical structures
Solution Approach 2:
The patent changes the feature representation parameters by using deep neural networks to learn optimal feature transformations from raw images. Instead of fixed handcrafted features, the system learns adaptive feature parameters that capture complex anatomical variations through multiple layers of non-linear transformations
2Device complexity
If handcrafted image features are used for anatomical object detection, then the computational process is simpler, but the computational complexity increases for large 2D+time or 3D volumetric images
Solution Approach 1:
The patent segments the computational process into hierarchical stages: first learning low-level features, then progressively learning higher-level abstract features through multiple neural network layers. This segmentation allows efficient processing by breaking down the complex task of detecting anatomical objects in large 3D volumes into manageable feature learning stages
Solution Approach 2:
The patent transitions from 2D image processing to 3D volumetric processing by extending the deep neural network architecture to handle three-dimensional medical images. The network learns features across all three spatial dimensions, enabling efficient detection in volumetric data without the computational burden of traditional 3D feature extraction methods
3Measurement precision
If deep neural networks are trained directly on image data to learn high-level features, then the detection accuracy improves for challenging cases, but the training and computational complexity increases
Solution Approach 1:
The patent applies preliminary unsupervised pre-training to initialize the deep neural network weights before supervised fine-tuning. This two-stage training approach allows the network to first learn general image features without labels, then specialize for anatomical detection with annotated data, improving convergence and accuracy while managing training complexity
Solution Approach 2:
The patent introduces intermediate feature representations at each layer of the deep neural network that serve as mediators between raw image data and final detection outputs. These intermediate features capture progressive levels of abstraction, allowing the network to build complex detection capabilities through composition of simpler feature detectors
Data Source
AI summary
A method and system for anatomical object detection using marginal space deep neural networks is disclosed. The pose parameter space for an anatomical object is divided into a series of marginal search spaces with increasing dimensionality. A respective sparse deep neural network is trained for each of the marginal search spaces, resulting in a series of trained sparse deep neural networks. Each of the trained sparse deep neural networks is trained by injecting sparsity into a deep neural network by removing filter weights of the deep neural network.