Deep Neural Network Object Identification in Medical Images
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep neural networks require large and high-resolution training sets, which are constrained by the nature of medical images, often coarse and limited in number, due to privacy concerns, making it difficult to leverage their features for object detection in medical imagery.
Innovation Solution
A computer-implemented method using a convolutional neural network (CNN) model-based system that includes a main deep neural network for identifying an object's approximate position and secondary deep neural networks for refining the location, with a statistical analysis stage to determine the object's position, and a training system that utilizes data augmentation tools to enhance the network's ability to identify objects across varying sizes and orientations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep neural networks are used for object identification, then identification accuracy and adaptability improve, but the requirement for large training datasets increases
Solution Approach 1:
The system performs preliminary actions by pre-training the deep neural network on large-scale natural images (ImageNet) before fine-tuning on medical images. This preliminary training establishes foundational feature extraction capabilities that transfer to medical image analysis, reducing the need for extensive medical image training data while maintaining high identification accuracy.
Solution Approach 2:
The patent applies transfer learning by copying knowledge and features learned from natural image datasets to medical image analysis. The pre-trained network's weights and feature extraction mechanisms are transferred and adapted to the medical domain, allowing the system to achieve accurate object identification with limited medical training data.
2Ease of manufacture
If conventional image processing algorithms are used, then implementation simplicity is maintained, but identification capability is limited to certain orientations and sizes
Solution Approach 1:
The system changes the fundamental parameters of the approach by transitioning from hand-crafted template matching to deep neural network-based feature learning. This parameter change enables the system to automatically adapt to various object sizes, orientations, and resolutions without requiring multiple fixed templates, significantly improving versatility while maintaining implementation feasibility through established deep learning frameworks.
Solution Approach 2:
The patent introduces a new dimension of abstraction by using hierarchical feature representations in deep neural networks. Instead of operating directly on pixel-level template matching, the system learns multi-scale feature representations across multiple network layers, enabling robust identification of objects regardless of size or orientation variations in the input medical images.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Methods, devices, and computer-readable storage media for identifying an object in an image, the method including using a first neural network to identify an approximate position of an object of interest in an image and identifying based on the approximate position, a section of the image that includes the object of interest. A plurality of sub-images corresponding to the identified section of the image are applied to a plurality of second neural networks to determine a plurality of second positions of the object of interest. The plurality of second positions are statistically analyzed to determine an output position of the object of interest.