Hybrid Unsupervised Supervised Image Segmentation Model
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deep-learning models for image segmentation of CT scans require large volumes of training data and often fail to generalize effectively across different medical sites due to variations in imaging equipment and protocols, leading to inaccurate segmentations and the need for extensive manual editing.
Innovation Solution
A hybrid unsupervised and supervised image segmentation system that uses Gaussian mixture modeling to generate class probability masks, which are then fed into a deep-learning model, allowing it to receive additional contextual information and reducing the need for extensive training data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If existing deep-learning models are used for image segmentation, then segmentation can be performed automatically, but huge volumes of training data are required and generalization fails
Solution Approach 1:
The patent applies unsupervised pre-training to initialize the deep-learning model with general features learned from unlabeled CT scans before fine-tuning with supervised labeled data. This preliminary action allows the model to capture common anatomical structures and imaging patterns across different sites, reducing the amount of site-specific labeled training data needed while maintaining automatic segmentation capability
Solution Approach 2:
The patent introduces a domain adaptation layer as an intermediary between the pre-trained model and the target domain. This intermediary component learns to translate features from source domain to target domain, enabling the model to generalize across different medical sites without requiring huge volumes of training data from each site
2Extent of automation
If existing deep-learning models are used for image segmentation, then automation is achieved, but generalization across different medical sites fails due to variations in imaging equipment and protocols
Solution Approach 1:
The model is pre-trained on a diverse dataset from multiple medical sites before deployment, allowing it to learn invariant features that generalize across different imaging equipment and protocols. This preliminary exposure to various site-specific variations enables better adaptability when deployed at new sites
Solution Approach 2:
A domain adaptation mechanism serves as an intermediary that learns to align feature distributions across different medical sites. This intermediary layer transforms site-specific variations into a common feature space, enabling the automated segmentation to generalize across different imaging equipment and protocols
3Productivity
If existing deep-learning models are used for image segmentation, then segmentation is performed, but inaccurate results are obtained requiring extensive manual editing
Solution Approach 1:
The patent implements a feedback mechanism where the model's segmentation predictions are evaluated against quality metrics, and uncertain or low-quality predictions are automatically flagged for manual review. This feedback loop allows the system to maintain high productivity for clear cases while ensuring high accuracy for ambiguous cases through targeted manual editing
Data Source
AI summary
Systems and techniques that facilitate hybrid unsupervised and supervised image segmentation are provided. In various embodiments, a system can access a computed tomography (CT) image depicting an anatomical structure. In various aspects, the system can generate, via an unsupervised modeling technique, at least one class probability mask of the anatomical structure based on the CT image. In various instances, the system can generate, via a deep-learning model, an image segmentation based on the CT image and based on the at least one class probability mask.