Bone Density Prediction Training for Consistent Planar Image Outputs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current AI models for predicting bone density exhibit high variability in output values for similar inputs, making them unsuitable for continuous monitoring due to the hard-to-interpret nature of AI models and the difficulty in controlling these variations.
Innovation Solution
Introduce a secondary loss term during model training using unlabeled image pairs with similar features and apply data augmentation techniques to enhance consistency, incorporating a precision term in the loss function and utilizing ROI extraction to focus on key features.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Difficulty of detecting and measuring
If AI models are used to predict bone density values, then screening capability is improved, but measurement consistency deteriorates
Solution Approach 1:
The training process is segmented into two distinct stages: pre-training on labeled data to learn basic bone density prediction, and subsequent fine-tuning on unlabeled image pairs to learn temporal consistency. This segmentation allows the model to acquire different skills in sequence, resolving the contradiction between detection capability and measurement consistency.
Solution Approach 2:
The model performs preliminary learning on labeled data to establish basic prediction capability, then prepares for consistency training by introducing unlabeled image pairs. This preliminary action on structured data before tackling the consistency problem enables the model to achieve both detection accuracy and measurement stability.
2Ease of manufacture
If conventional loss functions are used for training, then model training simplicity is maintained, but prediction consistency deteriorates
Solution Approach 1:
The loss function parameters are changed by introducing a temporal consistency term that operates on unlabeled data. This parameter modification allows the model to optimize for both accuracy and consistency without fundamentally changing the training architecture, maintaining relative simplicity while improving prediction stability.
Solution Approach 2:
Unlabeled image pairs serve as an intermediary element that bridges the gap between simple conventional training and consistent predictions. By introducing this intermediate training material with a specialized loss function, the model can gradually learn temporal consistency without completely redesigning the training pipeline.
3Quantity of substance
If AI models are trained on labeled data only, then training data requirements are reduced, but measurement reliability deteriorates
Solution Approach 1:
The model creates virtual training samples by using unlabeled image pairs as copies of real clinical data. These synthetic training examples allow the model to learn from abundant unlabeled data, effectively copying the structure and relationships present in labeled data without requiring additional manual annotations, thus improving reliability while maintaining data efficiency.
Data Source
AI summary
The present invention relates to a method of training a prediction model to generate a main predicted value of a main feature of an input image. The method comprises training the prediction model with a primary dataset containing labeled training images labeled with ground truth values and a secondary dataset containing two unlabeled training images without ground truth values. The training goal is to reduce both a first loss and a second loss, wherein the first loss calculates the difference between predicted values of the labeled training image and the ground truth values, and the second loss calculates the difference between predicted values of the two unlabeled training images.