Radiology Report Co-Training for Reliable Medical Image Pseudo-Labels
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The challenge of supervised learning for medical imaging using machine learning algorithms is the lack of annotated images, which limits model generalizability, and self-training methods amplify errors due to the model's inability to correct its own mistakes.
Innovation Solution
A co-training paradigm that leverages radiology reports to generate pseudo-labels for medical images, using a convolutional neural network for image classification and a radiology report classifier to improve training accuracy by exchanging information and reinforcing high-confidence labels.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If self-training method is used to utilize large amounts of unlabeled images, then the model can be trained with more data, but the model's errors are amplified because it cannot correct its own mistakes
Solution Approach 1:
The patent introduces radiology reports as an intermediary source of information to verify and correct pseudo-labels generated by the model. The reports act as a mediator between the model's predictions and the true labels, allowing the system to identify and correct errors without requiring manual annotation of all images.
Solution Approach 2:
The system implements feedback by using high-confidence pseudo-labels from the model to train the radiology report classifier, and then using the report classifier's predictions to verify and correct the model's pseudo-labels. This feedback loop allows the system to iteratively improve accuracy while utilizing unlabeled data.
2Productivity
If pseudo-labels are assigned based on model predictions with confidence threshold, then training can proceed with unlabeled data, but erroneous confident predictions are incorporated and reinforce errors
Solution Approach 1:
The radiology report classifier serves as an intermediary verification mechanism that checks the reliability of pseudo-labels before they are incorporated into training. Even when the model is confident in its predictions, the report classifier provides an independent verification layer to filter out erroneous labels.
Solution Approach 2:
The system applies partial verification by using confidence thresholds to selectively process only high-confidence predictions, rather than verifying all predictions equally. This partial action approach maintains training efficiency while still providing error correction for the most reliable pseudo-labels.
3Reliability
If manual annotation is performed to ensure label accuracy, then training data quality is high, but the lack of annotated images limits model generalizability
Solution Approach 1:
The system enables self-service by allowing the model and radiology report classifier to automatically generate and verify labels without requiring manual annotation. The model generates pseudo-labels for unlabeled images, and the report classifier verifies them, creating a self-sufficient labeling system that scales to large datasets.
Solution Approach 2:
Radiology reports serve as an intermediary that bridges the gap between manual annotation quality and automated labeling quantity. By leveraging the structured information in reports, the system achieves high label quality comparable to manual annotation while processing large volumes of unlabeled images automatically.
Data Source
AI summary
A method (100) of training a machine-learned (ML) image classifier (14) to classify images (30) respective to a set of labels includes: generating image-based labels for the images from the set of labels and image-based label confidence values for the image-based labels by applying the ML image classifier to the images; generating report-based labels for the images from the set of labels and report-based label confidence values for the report-based labels by applying a report classifier (16) to corresponding radiology reports; selecting a training subset of the set of images based on the image-based labels, the report-based labels, the image-based label confidence values, and the report-based label confidence values; assigning a pseudo-label for each image of the training subset which is one of the image-based label or the report-based label for the image; and training the ML image classifier using at least the selected training subset and the assigned pseudo-labels.