Global-Local Masked Autoencoder for Volumetric Medical Image Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing self-supervised learning methods for medical image analysis, such as Masked Autoencoder (MAE), face challenges in capturing global context information from volumetric medical images and ensuring stability of learned representations under input distortions, leading to slow convergence and inefficiency in training.
Innovation Solution
The Global-Local Masked AutoEncoder (GL-MAE) system addresses these challenges by incorporating global context learning through global-to-global and global-to-local consistency, reconstructing both global and local views, and using a momentum-based update mechanism to stabilize the learning process.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If local cropping strategies are used to reduce computation cost, then computational efficiency is improved, but global context information is lost
Solution Approach 1:
The patent divides the volumetric medical image into multiple local sub-volumes for parallel processing, enabling efficient computation while maintaining the ability to reconstruct global context through the autoencoder architecture that processes each sub-volume and integrates their representations
Solution Approach 2:
The patent employs a hierarchical structure where local sub-volume processing is nested within the global volume context, allowing the model to learn both local details and global relationships by embedding local representations within the overall volumetric structure
2Extent of automation
If strong augmentation with high mask ratio is used, then label efficiency is improved, but convergence speed deteriorates
Solution Approach 1:
The patent applies partial masking rather than complete or excessive masking, using a moderate mask ratio that provides sufficient self-supervised learning signal while preserving enough visible information to maintain stable gradient flow and fast convergence during training
Solution Approach 2:
The patent optimizes the mask ratio parameter to balance between label efficiency and convergence speed, finding the optimal masking level that maximizes self-supervised learning benefits without causing training instability or slow convergence
3Device complexity
If local sub-volumes are used for reconstruction, then computation cost is reduced, but representation stability deteriorates
Solution Approach 1:
The patent implements a feedback mechanism where the autoencoder reconstructs the original volumetric image from encoded representations of local sub-volumes, using the reconstruction loss to guide and stabilize the learning of robust representations that are invariant to the local processing distortions
Solution Approach 2:
The patent introduces an intermediary autoencoder architecture that mediates between local sub-volume processing and global representation stability, translating local features into stable global representations through the encoder-decoder structure that preserves essential volumetric relationships
Data Source
AI summary
A system is provided for training a neural network model. The system comprises one or more image acquisition devices, configured to obtain volumetric image data; and one or more processors, configured to: generate images of a plurality of views, which comprise a global complete view and one or more masked views; generate, by a Global-Local Masked AutoEncoder (GL-MAE) encoder system, representations corresponding to the global complete view and the one or more masked views; generating one or more reconstructed images corresponding to the one or more masked views; evaluating in a shared representation space, consistency between the representations of the global complete view and the one or more masked views; computing one or more losses based on the one or more reconstructed images and the results from the consistency evaluation; and updating one or more parameters in the neural network model based on the one or more losses.