Bayesian Target-Region Prediction With Uncertainty-Guided Label Screening

The patent implements a feedback mechanism where uncertainty information from the predictive model is fed back into the system to guide active learning. The system calculates uncertainty data (aleatoric and epistemic uncertainty) from predicted results and uses this feedback to screen and select medical images for labeling, which are then used to retrain and improve the predictive model, creating a continuous improvement loop that enhances reliability over time.

The system performs self-service by automatically screening medical images based on calculated uncertainty data and selecting images that would be most beneficial for model training. Instead of requiring manual selection of training images, the system autonomously identifies which images have high epistemic uncertainty and would provide the most value for improving the model, thereby serving itself in the active learning process.

The patent introduces an intermediary system that acts as a bridge between automated AI prediction and expert verification. The predictive model with uncertainty calculation serves as an intermediary that pre-screens and prioritizes images, allowing specialists to focus their expertise on cases with high uncertainty or potential abnormalities rather than reviewing every image, thus maintaining diagnostic accuracy while improving productivity.

Instead of requiring full expert review of all medical images, the system applies partial action by using the automated predictive model to handle the majority of routine cases with high confidence. Experts only need to perform excessive action (detailed review) on the small subset of images that the system identifies as having high uncertainty or potential issues, thereby maintaining diagnostic accuracy while significantly reducing the overall time and resource investment required.

The system performs preliminary action by initially deploying a predictive model trained with a small number of training data to quickly start the active learning process. This preliminary model, though not highly accurate initially, is sufficient to generate predictions and calculate uncertainty data, which then guide the selection of images for labeling and subsequent retraining, enabling the model to progressively improve accuracy over time.

The patent ensures continuity of useful action by implementing a continuous active learning loop where the predictive model continuously generates predictions, uncertainty is continuously calculated, images are continuously screened and selected for labeling, and the model is continuously retrained with new labeled data. This continuous cycle maintains and progressively improves measurement precision over time, transforming the initial limitation of small training data into a strength through ongoing learning.