Deep active learning framework for chest-abdominal CT scans segmentation

• Published in: Expert systems with applications 2025-03, Vol.263, p.125522, Article 125522
• Main Authors: Rokach, Lital, Aperstein, Yehudit, Akselrod-Ballin, Ayelet
• Format: Article
• Language: English
• Subjects: Active learning; Artificial intelligence; CT segmentation; Neural networks; Organ segmentation; Semi-supervised learning
• ISSN: 0957-4174
• DOI: 10.1016/j.eswa.2024.125522

State-of-the-art deep learning approaches for segmentation require large, high-quality annotated datasets for training and evaluation. However, the manual processes for creating these datasets are time-consuming, labor-intensive, and especially costly in medical imaging. To address this, we utilize an active learning (AL) strategy leveraging the power of machine learning to actively select informative samples from the unlabeled data for annotation. Our contributions include the creation and automation of a deep learning flow, applied to solve organ segmentation tasks in chest-abdominal CT scans leveraging AL. We have established an AL framework that automates the training of semantic segmentation models by prioritizing the labeling of images that are most valuable for the segmentation task. Our solution, designed for specific anatomical structures, has been further evaluated by incorporating additional organs, datasets, and AL acquisition functions. This adaptable framework can accommodate new organ segmentation tasks, different anatomies, and varied parameters within the AL process while reducing the annotation burden in medical imaging and allowing generalization to other application domains. We demonstrate the tangible benefits of the AL approach in terms of reduction in annotation cost and training time by the AL flow for a given supervised learning medical segmentation task. Our method achieved segmentation results for spleen (Dice score 0.908 [0.895,0.921], mIoU 0.833 [0.816, 0.849]) and liver (Dice score 0.855 [0.848, 0.861], mIoU 0.785 [0.777, 0.792]) using 2.5% of data available for training.