Neural Network Segmentation Accuracy via Radiologist Feedback
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current deep learning systems for brain tumor segmentation rely heavily on front-loaded human expertise for training and validation, lacking a complementary approach to back-load human expertise for evaluating and improving segmentation results.
Innovation Solution
A computer-implemented method and system that determines the accuracy of neural networks in producing computerized segmentations within MR images by saving segmented images on a network-connected computer, communicating with test computers to display images alongside manual segmentations for scoring, and receiving scores to update neural network configuration parameters.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If deep learning systems are used for automatic segmentation, then productivity is improved, but measurement precision deteriorates due to lack of human expertise validation
Solution Approach 1:
The system implements a feedback loop where radiologists score automated segmentations, and these scores are used to retrain and improve the deep learning model. This continuous feedback mechanism allows the system to maintain high productivity while progressively improving measurement precision through iterative learning from human expert evaluations.
Solution Approach 2:
The system performs preliminary automated segmentation before human review, allowing radiologists to focus their expertise on evaluating and correcting rather than creating segmentations from scratch. This preliminary action maintains productivity while the subsequent human review ensures measurement precision.
2Manufacturing precision
If human expertise is front-loaded for training, then manufacturing precision is improved, but loss of time increases due to extensive manual annotation
Solution Approach 1:
The system creates synthetic training data by copying and adapting existing manual segmentations through augmentation techniques, reducing the need for extensive new manual annotations while maintaining training data quality. This allows the system to improve manufacturing precision without proportionally increasing time investment.
Solution Approach 2:
The deep learning model performs self-training by automatically generating segmentations that are then scored and used to retrain the model. This self-service mechanism reduces dependency on continuous manual annotation, decreasing time loss while maintaining or improving training data quality through iterative self-improvement.
3Ease of operation
If automated segmentation is implemented, then ease of operation is improved, but reliability deteriorates due to black box nature of deep learning
Solution Approach 1:
The system enhances reliability by implementing feedback from radiologist scoring of automated segmentations. This external validation feedback loop provides confidence in the automated system's outputs while maintaining ease of operation, as the feedback mechanism systematically addresses reliability concerns without requiring manual intervention in every case.
Solution Approach 2:
The system introduces an intermediary validation layer where radiologists score segmentations before final acceptance. This intermediary step bridges the gap between automated operation and reliable validation, maintaining ease of operation while improving reliability through expert oversight of critical results.
Data Source
AI summary
The first objective comparison of automated and human segmentation of magnetic resonance images, or MRI, using a blinded controlled assessment study. Computers connected over a network divide duties including computerized segmenting of the images, manual segmenting of the images, comparison of the computer segmented images and the manually segmented images, and scoring of the images for accuracy. The scores are evaluated to update configuration parameters of a neural network.