Iterative multi-path tracking for video and volume segmentation with sparse point supervision.

Area of Science:

• Computer Vision
• Machine Learning
• Medical Imaging

Background:

• Supervised segmentation models require extensive pixel-wise annotations, which are costly and time-consuming to acquire.
• This is especially true for video and 3D volumetric data, posing a significant challenge for dataset aggregation.

Purpose of the Study:

• To develop a novel framework for producing precise pixel-wise segmentations with minimal supervision.
• To reduce the annotation burden for image segmentation tasks, particularly in video and volumetric data.

Main Methods:

• The framework utilizes 2D point supervision, requiring only a single 2D location per object per image.
• It employs semi-supervised learning to estimate object appearance and a graph-based optimization using a K-shortest path approach for segmentation.
• Iterative refinement of the object model and segmentation further enhances accuracy.

Main Results:

• The method achieves state-of-the-art segmentation performance across various objects and image modalities (video, 3D volumes).
• Utilizing gaze trackers for 2D point collection demonstrates practical efficacy.
• Generated segmentations can effectively train supervised machine learning classifiers.

Conclusions:

• This minimal supervision approach significantly lowers the cost and effort associated with creating high-quality image segmentations.
• The framework offers a viable solution for large-scale dataset creation in computer vision and medical imaging.
• The resulting segmentations are suitable for downstream machine learning tasks.