Learning Image Quality Assessment by Reinforcing Task Amenable Data Selection

In this paper, we consider a type of image quality assessment (IQA) as task-specific measurement, which can be used to select images that are more amenable given target task, such classification or segmentation. We propose train simultaneously two neural networks for selection and task using reinforcement learning. A controller network learns an policy by maximising accumulated reward based on the performance controller-selected validation set, whilst predictor is optimised training set. The trained therefore able reject lead poor accuracy in task. work, show controller-predicted IQA significantly different from labels manually defined humans. Furthermore, demonstrate it possible learn effective without “clean” thereby avoiding requirement human amenability. Using 6712, labelled segmented, clinical ultrasound 259 patients, experimental results holdout data proposed achieved mean $$0.94\pm 0.01$$ segmentation Dice $$0.89\pm 0.02$$ , discarding $$5\%$$ $$15\%$$ acquired images, respectively. improved was observed both tested tasks, compared with respective $$0.90\pm $$0.82\pm considering This enables feedback during real-time acquisition among many other medical imaging applications.