A Deep Active Learning Framework with Information Guided Label Generation for Medical Image Segmentation

Acquiring medical images and their segmentation labels is often time-consuming and labor-intensive. Compared to natural images, medical images need to be screened and annotated by professional doctors in segmentation tasks, especially those containing multiple organ tissues. To reduce the workload of doctors, we propose a deep active learning-based framework for medical image segmentation. Instead of all, the proposed framework can select an optimal number of medical images for label generation by doctors. The selected images are enough to train a good segmentation model because they are diversified, informational, and unique to represent the whole dataset. The proposed method consists of three steps: (1) Using an auto-encoder to extract features from unlabeled images and applying feature clustering and image entropy to select an initial subset of images for segmentation label generation; (2) Employing the mean-teacher method to train a segmentation model in a semi-supervised manner; (3) Updating the labeled subset and reducing redundancy by a deduplication uncertainty query strategy. Extensive experiments are conducted on a breast ultrasound dataset and a knee cartilage ultrasound dataset to evaluate the performance of the proposed method. Experimental results prove that our deep active learning framework can significantly reduce the number of labeled samples while achieving comparable segmentation results to fully-labeled supervision.