MAPLE - Manifold-Aware Point-Lifted evolution for very weakly supervised self learning in medical image segmentation
Year: 2026
Type: article
Abstract: ABSTRACT Accurate image segmentation remains a central problem in pattern recognition, particularly under limited supervision. While fully supervised convolutional and vision–language models achieve strong performance, they rely on dense annotations and multimodal supervision, restricting scalability in label-scarce settings. We propose MAPLE (Manifold-Aware Point-Lifted Evolution), a weakly supervised segmentation framework that requires only sparse point-level annotations. MAPLE formulates segmentation as a manifold-guided self-learning process in which foreground and background points are propagated along intrinsic image structures to generate conservative pseudo-labels. These pseudo-labels are iteratively refined through point-lifted localization learning, pseudo self-refinement, and uncertainty-aware consistency distillation within a lightweight 2.5D attention-based U-Net architecture. The framework prioritizes structure-aware propagation and learning dynamics over architectural complexity or pretrained priors. Extensive evaluation on six multi-organ datasets demonstrates that MAPLE achieves competitive Dice similarity compared to fully supervised convolutional models. Although a performance gap remains relative to supervised vision–language models, boundary-aware and size-sensitive metrics reveal minimal practical discrepancies in several datasets. These results indicate that large gains in overlap-based metrics do not necessarily imply proportional improvements in task-relevant quantification, underscoring the importance of evaluation criteria aligned with downstream objectives.
Source: Pattern Recognition
Institution: The Ohio State University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $2,590
OpenAlex ID: W7208822046