Active learning using adaptable task-based prioritisation.

Area of Science:

• Medical Image Computing
• Artificial Intelligence
• Machine Learning

Background:

• Supervised machine learning for medical imaging requires extensive expert annotation, which is time-consuming and costly.
• Unlabeled medical image data is often abundant, presenting an opportunity for efficient learning strategies.

Purpose of the Study:

• To develop an adaptable active learning strategy for label-efficient medical image segmentation.
• To create a controller neural network that prioritizes images for expert annotation in batch-mode active learning.

Main Methods:

• A controller neural network was developed to measure image priority within batches for multi-class segmentation.
• The controller was optimized using a meta-reinforcement learning algorithm within a Markov decision process (MDP) framework.
• The approach was validated using CT datasets from over a thousand patients across nine abdominal organ segmentation tasks.

Main Results:

• The proposed adaptable prioritization metric achieved converging segmentation accuracy for a novel kidney segmentation task using 40-60% fewer labels compared to heuristic or random methods.
• Significant performance improvements were observed: 22.6% in Dice score for kidney segmentation and 10.2% for liver vessel segmentation compared to random prioritization.
• The controller demonstrated cross-institute and cross-organ adaptability, effectively prioritizing images for new segmentation tasks.

Conclusions:

• The developed meta-reinforcement learning approach enables an adaptable prioritization controller for efficient medical image annotation.
• This method substantially reduces the required amount of labeled data for training accurate medical image segmentation models.
• The adaptable prioritization strategy shows strong potential for improving the efficiency and effectiveness of machine learning in clinical settings.