Keya Medical’s computer scientists, working with colleagues at the Shenzhen Second People’s Hospital, have devised a new method that significantly reduces the amount of annotation needed for accurate image segmentation.
This study “Annotation-Efficient COVID-19 Pneumonia Lesion Segmentation using Error-Aware Unified
Semi-supervised and Active Learning,” published Feb 22, 2022, in the Journal of IEEE Transactions on Artificial Intelligence, describes an Error-Aware Unified Semi-supervised and Active Learning approach (EA-SSAL). This study focused on COVID-19 pneumonia lesion segmentation in 3D Computed Tomography (CT) images.
This new approach uses an error-aware, unified Semi-Supervised Learning SSL and Active Learning (AL) segmentation framework. This annotation-efficient framework makes full use of existing unlabeled data and proactively identifies informative images to annotate next.
This improved annotation-efficient segmentation approach has three advantages over conventional methods:
(a) A novel feed-forward error estimator network is proposed to estimate potential errors made by the current segmentation network, which will guide learning from the unlabeled images, as well as identify informative images/regions to annotate next.
(b) The independent error estimator can capture both confident and uncertain errors made by the segmentation network.
(c) The same error estimator unifies semi-supervised learning and active learning, without additional elaborate training or computing. Experiments show improving performance over using either technique alone.
The article provides details on how this method can generate high-quality image segmentation results using fewer human annotation resources than required in conventional approaches. The core idea presented is a unified SSL-AI method that achieves annotation-efficient segmentation by being error-aware.
For article access see https://ieeexplore.ieee.org/document/9699409.
About Keya Medical
Keya Medical is an international medical technology company developing deep learning-based medical devices for disease diagnosis and treatment. The company is committed to creating solutions that deliver clinical value at all stages in the patient care process, covering specialties including cardiology, neurology, pulmonology, pathology, and surgery. Since 2016, Keya Medical has collaborated with more than 725 hospitals to improve diagnostics, care processes, and clinical outcomes. To learn more about Keya Medical, follow us on LinkedIn, Twitter, and Facebook.
Recent Comments