Eigenrank by committee: Von-Neumann entropy based data subset selection and failure prediction for deep learning...

Area of Science:

• Medical image analysis
• Artificial intelligence in healthcare
• Computational anatomy

Background:

• Manual segmentation of medical images is crucial for training deep learning (DL) algorithms but is time-consuming and costly.
• This limits the use of vast imaging datasets, hindering the development of accurate and robust segmentation models.
• Efficient data subset selection is vital for maximizing the utility of available medical imaging data.

Purpose of the Study:

• To introduce a novel algorithm, Eigenrank by Committee (EBC), for intelligent selection of data subsets for training DL models.
• To evaluate the effectiveness of EBC in improving the accuracy and robustness of medical image segmentation compared to random sampling.
• To address the challenge of selecting optimal training data from large medical imaging databases.

Main Methods:

• The Eigenrank by Committee (EBC) algorithm identifies informative data points by maximizing disagreement among DL models in a committee.
• Disagreement is quantified using the maximum eigenvalue of a Dice coefficient disagreement matrix.
• EBC was used to select subsets for training U-Net models for spinal canal and intervertebral disk segmentation, compared against random sampling.

Main Results:

• Deep learning models trained on EBC-selected subsets achieved significantly higher average Dice coefficients on unseen data compared to those trained on randomly selected subsets (p < 0.05).
• Segmentations generated by models trained with EBC exhibited significantly lower variance in Dice coefficients (p < 0.05), indicating greater robustness.
• EBC demonstrated superior performance in selecting data subsets that lead to more accurate and consistent segmentation outcomes.

Conclusions:

• The Eigenrank by Committee (EBC) algorithm offers an effective strategy for selecting optimal training data subsets in medical image segmentation.
• EBC-selected data subsets lead to more accurate and robust deep learning models compared to random sampling.
• This approach can enhance the development of AI-driven medical imaging tools by maximizing the value of limited manual annotation efforts.