Parallel 2D Branches for Pneumonia Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current automatic screening methods for pneumonia using CT images face challenges such as loss of original information due to varying image sequences, inaccurate classification results affecting subsequent diagnoses, and high computational demands of 3D deep learning networks, which lead to decreased diagnostic accuracy.
Innovation Solution
The proposed solution involves a deep learning network with three parallel 2D classification branches processing transverse, sagittal, and coronal images, utilizing convolutional neural networks and recurrent neural networks to extract features and fuse information across different layers, reducing computational load and improving diagnostic precision.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a 3D convolutional neural network is used for pneumonia diagnosis, then the diagnostic accuracy can be improved by utilizing 3D spatial information, but the computational load, video memory consumption, and training time increase significantly
Solution Approach 1:
The patent segments the 3D CT image into multiple 2D slices along different directions (transverse, sagittal, coronal views). Each 2D slice is processed independently through separate 2D convolutional neural networks, and the results are fused to achieve comprehensive 3D analysis. This segmentation approach maintains diagnostic accuracy by preserving spatial information while significantly reducing computational complexity compared to processing the entire 3D volume at once.
Solution Approach 2:
The patent transforms the 3D diagnosis problem into a multi-dimensional 2D analysis by extracting features from transverse, sagittal, and coronal views. Instead of using a single 3D convolutional network, the system processes images in multiple 2D dimensions and fuses the results, achieving equivalent or superior diagnostic accuracy with lower computational requirements.
2Device complexity
If n images are selected from a sequence of 2D CT images for classification, then the computational complexity is reduced, but original information of the pneumonia lesion is lost
Solution Approach 1:
Instead of selecting only n images from the CT sequence, the patent divides the entire 3D CT volume into multiple 2D slices and processes all of them through the neural network. The 3D spatial relationships between slices are preserved and utilized through the multi-directional view fusion, ensuring no lesion information is lost while maintaining manageable computational complexity.
Solution Approach 2:
The patent merges information from multiple 2D slices and multiple viewing directions (transverse, sagittal, coronal) to form a comprehensive diagnosis. By combining all slice information rather than selecting a subset, the system preserves complete lesion information while using efficient 2D processing and fusion strategies to control computational complexity.
3Device complexity
If a series model structure is used where the first classification result affects subsequent determinations, then the processing flow is simplified, but diagnostic accuracy decreases when the first classification is inaccurate
Solution Approach 1:
The patent segments the diagnostic process into parallel independent classification branches for different viewing directions (transverse, sagittal, coronal) rather than using a sequential series structure. Each branch independently processes its 2D image sequence and produces a classification result, eliminating the propagation of errors that occurs in series models where one classification affects subsequent determinations.
Solution Approach 2:
The patent merges the classification results from multiple parallel branches through a fusion mechanism that aggregates information from all directions. This parallel-then-merge approach allows each branch to independently analyze its data without being influenced by errors in other branches, while the fusion step combines all perspectives to achieve the final diagnosis, improving robustness against individual classification errors.
Data Source
AI summary
The present disclosure provides a method, a device, and a non-transitory computer-readable storage medium for detecting a medical condition of an organ. The method includes obtaining 2D image sequences of the organ in a plurality of different directions and applying a plurality of classification branches to the 2D image sequences. Each classification branch receives a 2D image sequence of one direction and provides a classification result with respect to that direction. Each classification branch includes a convolutional neural network configured to extract first image features from the corresponding 2D image sequence and a recurrent neural network configured to extract second image features from the first image features. The method further includes fusing the classification results provided by the plurality of classification branches for detecting the medical condition.