Liver Segmentation Using Pre-trained Models and Graph Cut
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current liver segmentation methods in medical imaging are time-consuming, require user interaction, and struggle to accurately extract liver anatomy from normal CT scans without contrast, especially due to similarities in intensity with neighboring organs like the heart and kidney.
Innovation Solution
An automatic liver segmentation method using pre-trained models for liver, heart, and kidney regions, combined with a graph cut algorithm, to identify internal and external regions and produce a liver mask, which operates on both contrast and normal CT images, reducing user interaction and improving accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual delineation is used for liver segmentation, then accuracy can be maintained, but the process becomes time-consuming and challenging
Solution Approach 1:
The patent applies preliminary action by using pre-trained models to generate initial liver segmentation results before the final graph cut optimization. The pre-trained models (U-Net, V-Net, or Attention U-Net) perform initial segmentation to identify potential liver regions, which then serve as starting points for the subsequent graph cut algorithm to refine and finalize the segmentation boundaries, thereby reducing the overall processing time while maintaining accuracy
Solution Approach 2:
The patent introduces an intermediary approach by combining deep learning-based pre-trained models with the graph cut algorithm. The pre-trained model acts as an intermediary that provides initial segmentation estimates, which are then refined by the graph cut algorithm. This intermediary step allows the system to leverage the speed of automated modeling while maintaining the precision of iterative optimization methods
2Productivity
If automatic segmentation methods are used, then processing speed improves, but accuracy deteriorates due to anatomical shape variations
Solution Approach 1:
The patent applies dynamics by implementing an iterative graph cut algorithm that dynamically adjusts segmentation boundaries based on image gradients and regional statistics. Rather than using a static thresholding approach, the algorithm iteratively refines the segmentation by evaluating boundary energy and adjusting the分割 boundaries to better fit the actual liver anatomy, thereby improving accuracy while maintaining automated processing speed
Solution Approach 2:
The patent incorporates feedback mechanisms through the graph cut algorithm, which continuously evaluates the segmentation quality by computing boundary energy and regional homogeneity. The algorithm uses this feedback to iteratively adjust the segmentation boundaries, comparing the current segmentation against image features and refining the results until convergence, thus improving accuracy without sacrificing processing speed
3Object-affected harmful factors
If liver segmentation is performed on normal CT images without contrast, then patient safety improves, but segmentation accuracy deteriorates due to intensity similarities with neighboring organs
Solution Approach 1:
The patent replaces the mechanical/chemical approach of using contrast agents with a computational approach using graph cut algorithms and deep learning models. Instead of relying on contrast-enhanced intensity differences to distinguish liver from neighboring organs, the system uses learned features from training data and gradient-based boundary detection to identify liver boundaries in non-contrast images, thereby eliminating contrast exposure while maintaining segmentation accuracy
Solution Approach 2:
The patent applies parameter changes by transforming the segmentation problem from intensity-based thresholding to energy minimization based on boundary gradients and regional statistics. The graph cut algorithm changes the parameter space from simple pixel intensity values to composite energy functions that incorporate boundary strength, regional homogeneity, and spatial coherence, enabling accurate liver segmentation in non-contrast images where intensity differences are minimal
4Productivity
If pre-trained models are used for segmentation, then processing efficiency improves, but model complexity increases
Solution Approach 1:
The patent applies segmentation by dividing the overall segmentation task into distinct components: pre-trained model initialization (using U-Net, V-Net, or Attention U-Net architectures) and graph cut refinement. This segmentation of the processing pipeline allows each component to be optimized independently, managing complexity by breaking down the challenging liver segmentation task into more manageable sub-tasks that can be executed sequentially
Data Source
AI summary
A method for identifying a liver in a CT image of a patient is provided. The method includes applying a liver model to the CT image. The method further includes extracting an internal liver region and an external liver region from the CT image based on the applied liver model. The method also includes performing a graph cut algorithm on the CT image based on the internal liver region and the external liver region to produce a liver image. The performing of the graph cut algorithm on the CT image to produce the liver image may be further based on an internal heart and/or kidney region and an external heart and/or kidney region. A non-transitory computer-readable storage medium encoded with a program is provided.