Liver Fibrosis Recognition Model Using Hetero-Image Fusion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for liver fibrosis assessment using ultrasound images are invasive, subjective, and lack sensitivity and specificity, particularly when dealing with inflammation and liver steatosis, and existing automated solutions face challenges with variable image numbers and high computational costs.
Innovation Solution
A method involving segmenting interest regions in liver medical images, generating feature maps, and using global hetero-image fusion and view-specific parameterization to iteratively train a liver fibrosis recognition model, focusing on relevant clinical features and allowing for arbitrary numbers of images, thereby improving robustness and practicality.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If conventional ultrasound assessment is used for liver fibrosis, then versatility is improved, but measurement precision deteriorates due to subjective measurement and high inter- and intra-rater variability
Solution Approach 1:
The patent replaces the manual, subjective mechanical assessment process with an automated deep learning-based image processing system. The CNN model automatically extracts features and generates fibrosis scores, eliminating inter- and intra-rater variability while maintaining versatility across different ultrasound images and clinical settings.
Solution Approach 2:
The system enables self-service by allowing the ultrasound images to automatically generate fibrosis assessments without requiring manual annotation or expert interpretation for each case. The trained model independently processes images and produces consistent, objective measurements.
2Ease of manufacture
If elastography is used for liver fibrosis assessment, then non-invasive capability is improved, but reliability deteriorates due to confounding factors such as inflammation, liver steatosis, and patient etiology
Solution Approach 1:
The patent extracts and isolates the liver parenchyma region of interest from the full ultrasound images using segmentation techniques. By focusing specifically on the liver tissue and excluding surrounding structures and artifacts, the system eliminates confounding factors such as inflammation and steatosis that affect elastography, while maintaining the non-invasive advantage of ultrasound.
Solution Approach 2:
The deep learning model acts as an intermediary that processes the raw ultrasound images and extracts meaningful fibrosis indicators while filtering out confounding factors. The model learns to distinguish between fibrosis-related features and artifacts from inflammation or steatosis through training on labeled data.
3Measurement precision
If feature concatenation approach is used to fuse features from multiple ultrasound images, then prediction accuracy is improved, but computational cost and memory requirements drastically increase
Solution Approach 1:
The patent merges the processing of multiple ultrasound images into a unified deep learning framework where features from all images are processed simultaneously through shared convolutional layers. This approach fuses information from multiple images to improve prediction accuracy while avoiding the exponential computational cost of traditional feature concatenation methods.
Solution Approach 2:
The model employs universal, view-specific parameterization that allows the same network architecture to process any number of ultrasound images from different views. The system is designed to handle arbitrary numbers of input images without requiring separate processing pipelines, reducing computational overhead while maintaining the ability to fuse features from multiple sources.
4Measurement precision
If manually labeled indicators are used as ancillary supervision, then model training is improved, but labor costs and time consumption increase considerably
Solution Approach 1:
The system uses self-supervised learning where the model learns to generate fibrosis scores by processing the ultrasound images themselves without requiring extensive manual annotation. The training process leverages the inherent information in the images, reducing the need for time-consuming manual labeling while still achieving high training quality.
Data Source
AI summary
A liver fibrosis recognition method based on medical images and a computing device using thereof obtains a plurality of first binary images by segmenting a region of interest in each of a plurality of medical images of a liver. A rectangular region is created for each first binary image, and a plurality of second binary images is obtained by generating a second binary according to each rectangular region and the first binary image. A feature map is obtained from each liver medical image and images are generated according to the second binary images and corresponding to the plurality of feature maps. A model for recognition is iteratively trained based on the plurality of final images and recognition of liver fibrosis in patients is then achievable using the model.