3D Epileptogenic Focus Localization via SiameseNet PET Analysis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional medical imaging technologies for epilepsy, such as PET, face challenges in accurately and efficiently locating epileptogenic foci due to reliance on two-dimensional methods, sensitivity to registration errors, and limited data quality, leading to reduced detection sensitivity and high dependence on clinical experience.
Innovation Solution
A three-dimensional automatic location system based on deep learning, utilizing PET image registration and a SiameseNet architecture with data preprocessing and image block division, to enhance feature extraction and classification accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional visual evaluation methods are used for PET image analysis, then clinical experience can guide diagnosis, but the process becomes very time-consuming and detection sensitivity is reduced
Solution Approach 1:
The patent replaces the mechanical visual evaluation system with an automated computer-aided diagnosis system that uses image processing algorithms to analyze PET images. The system automatically extracts features, performs registration, and generates diagnostic results, substituting human visual inspection with computational analysis to achieve both high sensitivity and efficiency
Solution Approach 2:
The system enables self-service diagnosis by automating the entire analysis pipeline from image input to result generation. The computer-aided diagnosis system independently performs feature extraction, image registration, and diagnostic decision-making without requiring continuous human intervention, thereby reducing diagnosis time while maintaining detection sensitivity
2Ease of manufacture
If regional statistical methods are used to divide brain into larger regions of interest, then processing is simplified, but subtle changes are ignored resulting in reduced detection sensitivity
Solution Approach 1:
The patent applies segmentation by dividing the brain image into multiple regions of interest based on anatomical structures and functional areas. This allows the system to process complex brain images systematically while maintaining the ability to detect subtle changes in specific regions, balancing processing simplicity with detection sensitivity
Solution Approach 2:
The system implements local quality analysis by applying different processing strategies to different regions of the brain. Important regions with potential epileptogenic foci receive more detailed analysis with higher resolution processing, while less critical areas use simplified processing, thereby maintaining detection sensitivity for subtle changes while keeping overall processing manageable
3Measurement precision
If voxel statistical methods are used with SPM software to compare individual cases and control groups, then detailed analysis is achieved, but registration errors cause false positives in misaligned regions
Solution Approach 1:
The patent applies preliminary action by performing rigid and affine registration operations before the main statistical comparison. This pre-alignment ensures that images from different subjects are properly oriented and positioned relative to each other, eliminating registration errors that would otherwise cause false positives in the subsequent detailed voxel-wise statistical analysis
Solution Approach 2:
The system uses a standardized template brain image as an intermediary reference for all registrations. By comparing all individual images to this common template rather than directly to each other, the system ensures consistent alignment across all comparisons, reducing false positives while maintaining the ability to perform detailed voxel-level statistical analysis
4Adaptability or versatility
If two-dimensional location algorithms are used for PET image processing, then existing algorithms can be applied, but important inter-frame information is ignored
Solution Approach 1:
The patent transitions from two-dimensional image analysis to three-dimensional volumetric processing. The system processes the entire 3D PET volume, analyzing relationships across all spatial dimensions and utilizing inter-frame information. This dimensional expansion enables more accurate localization of epileptogenic foci by considering the full three-dimensional context rather than analyzing individual 2D slices in isolation
Data Source
AI summary
The present disclosure discloses a three-dimensional automatic location system for an epileptogenic focus based on deep learning. The system includes: a PET image acquisition and labelling module; a registration module mapping PET image to standard symmetrical brain template; a PET image preprocessing module generating mirror image pairs of left and right brain image blocks; a network SiameseNet training module containing two deep residual convolutional neural networks which share weight parameters, an output layer connecting a multilayer perceptron and a softmax layer, and using a training set of an epileptogenic focus image and an normal image to train the network to obtain a network model; a classification module and epileptogenic focus location module, using the trained network model to generate a probabilistic heatmap for the newly input PET image, a classifier determining whether the image is normal or epileptogenic focus sample, and then predicting a position for the epileptogenic focus region.