Liver CT Segmentation via Mixed Supervised Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing liver CT image segmentation methods face challenges in obtaining accurate pixel-level annotations, which are time-consuming and costly, and semi-supervised learning methods struggle with incorrect labels and limited precision improvement, while mixed supervision methods often result in inconsistent results due to the large number of independent parameters across tasks.
Innovation Solution
A liver CT image segmentation system based on mixed supervised learning, utilizing an image preprocessing unit, feature extraction unit, word vector segmentation unit, and single-layer convolutional classification unit, with a learnable word vector and triplet loss to enhance feature representation and reduce parameter inconsistency, allowing for efficient use of weak labels and improved segmentation precision.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If pixel-level annotation is performed to obtain strong labels for supervised learning, then segmentation precision is improved, but annotation time and manpower costs increase significantly
Solution Approach 1:
The patent applies partial action by using image-level annotations (weak labels) instead of complete pixel-level annotations (strong labels). The weak labels indicate only whether a liver is present in the image, which is much faster to annotate while still providing useful supervision signal for the segmentation task, thereby reducing annotation time while maintaining reasonable segmentation precision
Solution Approach 2:
The patent introduces an intermediary approach by using weak labels as a middle ground between no labels and strong labels. These weak labels serve as an intermediate supervision signal that is easier to obtain than strong labels but still more informative than no labels, enabling efficient training without requiring expensive pixel-level annotations
2Measurement precision
If pixel-level annotation is performed to obtain strong labels, then segmentation precision is improved, but annotation costs increase significantly
Solution Approach 1:
The patent uses partial action by implementing weak label annotation that requires only image-level information (presence/absence of liver) rather than complete pixel-level segmentation. This partial annotation approach dramatically reduces the time and expertise required, thereby lowering annotation costs while still providing effective supervision for training the segmentation model
3Adaptability or versatility
If data augmentation is used to expand training dataset, then generalization ability is improved, but effective data is wasted
Solution Approach 1:
The patent applies self-service by enabling the model to learn effectively from weak labels without requiring extensive data augmentation. The weak label supervision mechanism itself provides sufficient learning signal that allows the model to generalize well, thereby eliminating the need for resource-intensive data augmentation operations and avoiding waste of effective data
4Quantity of substance
If semi-supervised learning with pseudo labels is used, then training dataset is expanded, but incorrect labels affect network learning
Solution Approach 1:
The patent converts the potential harm of using weak labels (which are less informative than strong labels) into a benefit by designing a training framework that specifically leverages weak label supervision. The system is designed to extract maximum value from the limited information in weak labels, turning what could be a limitation into an advantage for efficient training
5Adaptability or versatility
If multi-task framework is used for mixed supervision, then generalization ability is enhanced, but parameter inconsistency increases
Solution Approach 1:
The patent applies merging by integrating the segmentation task and classification task into a unified framework where both tasks share the same network architecture and are trained jointly with a combined loss function. This merging approach ensures parameter consistency between tasks while still benefiting from the complementary supervision signals provided by strong and weak labels
Data Source
AI summary
A liver CT image segmentation system and algorithm based on mixed supervised learning is provided. The image segmentation system includes an image preprocessing unit, a feature extraction unit, a word vector segmentation unit and a single-layer convolutional classification unit. The image preprocessing unit is in data connection with the feature extraction unit. The feature extraction unit is respectively in data connection with the word vector segmentation unit and the single-layer convolutional classification unit. In the present disclosure, a multi-task framework is used for respectively performing segmentation and classification tasks, to achieve high segmentation precision through a large number of weak label data and a small number of strong labels.