Supervised machine learning-based medical image computing applications require expert label curation while unlabeled image data is relatively abundant. The active learning method aims to prioritize a subset of available image data for expert annotation for label-efficient model training. We develop a controller neural network that measures the priority of images within a set of batches, similar to batch-mode active learning, for multi-class segmentation tasks. The controller is optimized by rewarding task-specific performance gains within a Markov Decision Process (MDP) environment that also optimizes task predictors. In this study, the task predictor is a segmentation network. A meta-reinforcement learning algorithm using multiple MDPs has been proposed, in which a pre-trained controller can be used to train new controllers containing data from different institutions and/or requiring segmentation of different organs and structures within the abdomen. Can be adapted to MDP. We present experimental results using multiple CT datasets from over 1,000 patients with nine different abdominal organ segmentation tasks and demonstrate the effectiveness of the learned prioritization controller function and its cross-institutional and Demonstrate cross-organ adaptability. The proposed adaptive prioritization metric provides convergent segmentation for a new kidney segmentation task not seen in training, using approximately 40% to 60% of the labels required by other heuristic or random prioritization metrics. Indicates that it yields precision. For clinical datasets with limited size, the proposed adaptive prioritization improves Dice for kidney and liver vessel segmentation tasks, respectively, compared to random prioritization and alternative active sampling strategies. It yields a performance improvement of 22.6% and 10.2% in score.
