Sparse annotation learning for dense volumetric MR image segmentation with uncertainty estimation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Training neural networks for medical image segmentation demands extensive, accurately annotated data.
• Acquiring such detailed annotations is labor-intensive, time-consuming, and requires expert knowledge.
• This scarcity of annotated data presents a significant challenge in the medical imaging domain.

Purpose of the Study:

• To develop a neural network framework capable of performing 3D volumetric segmentation with extremely limited annotated data.
• To address the challenge of insufficient training samples in medical image segmentation.
• To explore methods for effectively utilizing unlabeled data in the segmentation process.

Main Methods:

• Proposed the extremely sparse annotation neural network (ESA-Net) framework for 3D segmentation.
• Developed a four-component architecture including intra-slice pixel dependency, inter-slice correlation, pseudo-label fusion, and network optimization modules.
• Employed uncertainty estimation, temporal ensembling, self-supervised registration, and rotation ensembling for label generation and propagation.

Main Results:

• ESA-Net demonstrated superior segmentation performance on challenging magnetic resonance image segmentation tasks.
• The framework consistently achieved better results compared to five state-of-the-art methods under extremely sparse annotation conditions.
• Validated the effectiveness of exploiting information from unlabeled data for improved segmentation.

Conclusions:

• ESA-Net offers an effective solution for 3D medical image segmentation with minimal annotations.
• The proposed method successfully leverages both intra-slice and inter-slice information from unlabeled data.
• Highlights the potential of uncertainty estimation and novel pseudo-labeling strategies in data-scarce segmentation scenarios.