Active relearning for robust supervised training of emphysema patterns.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pulmonology

Background:

• Computed tomography (CT) scans are crucial for identifying lung parenchymal abnormalities.
• Subjective interpretation in differential diagnosis of emphysema necessitates automated classification methods.
• Supervised learning for emphysema classification relies heavily on the quality of training samples.

Purpose of the Study:

• To develop and evaluate a physician-in-the-loop feedback approach for optimizing emphysema classification training samples.
• To minimize ambiguity in training data selection for improved automated diagnosis.
• To enhance the accuracy and robustness of emphysema classification on CT scans.

Main Methods:

• A radiologist selected regions of interest (ROIs) from 15 CT datasets for training.
• An ensemble of seven un-optimized support vector models (SVM) was constructed using multi-view inductive learning.
• Active relearning incorporated expert feedback to resolve label conflicts, refining the training set.

Main Results:

• The physician-in-the-loop feedback with un-optimized SVMs increased classification accuracy by 15% and reduced support vectors by 25%.
• An average accuracy improvement of 21% was observed across six general-purpose classifiers in their optimized parameter space.
• The cooperative feedback method demonstrated enhanced training sample quality for automated emphysema classification.

Conclusions:

• The proposed just-in-time active relearning feedback significantly improves automated emphysema classification accuracy on CT scans.
• This cooperative approach enhances training data quality, paving the way for more precise emphysema quantification.
• Physician-in-the-loop strategies are effective in refining machine learning models for medical image analysis.