Active Learning Clustering for Scan Study Mapping
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The lack of standardized naming for imaging procedures across different countries and institutions hinders comparison and establishment of national or international registries, and manual mapping of scan studies to standardized identifiers is cumbersome and time-consuming for radiologists.
Innovation Solution
A system and method for training machine learning systems using a minimal number of labels, employing a clustering module to classify scan studies, an active learning module to select and label key studies, and a machine learning system training module to map scan studies to standardized identifier codes, with an evaluation metric based on entropy to optimize labeling efforts.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual mapping of scan studies to standardized identifiers is performed, then accuracy of mapping can be maintained, but time consumption and labor effort increase significantly
Solution Approach 1:
A clustering module is introduced as an intermediary between manual labeling and final classification. The module automatically groups scan studies into clusters based on similarity, and only selects representative samples from each cluster for manual labeling. This intermediary step reduces the number of manual labeling tasks while maintaining mapping accuracy through the clustering structure.
Solution Approach 2:
The system enables self-service by allowing the machine learning model to automatically perform mapping tasks for most scan studies based on training from clustered samples. The model serves itself by generalizing from the clustered training data to classify new scan studies without requiring continuous manual intervention for every case.
2Measurement precision
If supervised learning with large labeled datasets is used to train machine learning systems, then model accuracy improves, but the cost and time of data preparation increases
Solution Approach 1:
The dataset is segmented into multiple clusters based on scan study characteristics. Instead of treating all data uniformly, the segmentation allows the system to select only representative samples from each cluster for labeling, reducing the total volume of labeled data needed while maintaining comprehensive coverage of the data space.
Solution Approach 2:
The system applies partial action by labeling only a subset of data points (cluster representatives) rather than the entire dataset. This partial labeling approach is sufficient to train the model effectively, avoiding the excessive effort of labeling all available data while still achieving good model accuracy.
3Adaptability or versatility
If standardized identifier codes are implemented across institutions, then comparability of imaging procedures improves, but implementation complexity and coordination requirements increase
Solution Approach 1:
The system creates a standardized mapping framework that can be copied and deployed across different institutions. The machine learning model and clustering approach serve as a reusable template that maintains comparability without requiring each institution to develop custom solutions, reducing implementation complexity through replication.
Data Source
AI summary
Active learning is used to control which scan studies are to be mapped by a user. This control is utilized to prompt for labeling of the relatively difficult data points for a machine learning system in its current state of training. A number of techniques of mining knowledge from the scan studies and for determining optimal decision criteria are also provided.