Conceitos Básicos
Achieving diversified and personalized results in multi-rater medical image segmentation through a two-stage framework.
Resumo
The study addresses the challenges of annotation ambiguity in medical image segmentation due to data uncertainties and observer preferences. It introduces a novel two-stage framework, D-Persona, focusing on diversification and personalization. Stage I constructs a common latent space for diverse expert opinions, while Stage II adapts attention-based projection heads for personalized segmentation. Extensive experiments demonstrate superior performance in providing diversified and personalized results simultaneously.
Directory:
Introduction
Importance of automatic medical image segmentation.
Challenges posed by annotation ambiguity.
Related Work
Overview of crowdsourcing, generation-based, and one-stage personalization methods.
Methods
Description of the two-stage D-Persona framework.
Stage I: Diversified Segmentation with bound-constrained loss.
Stage II: Personalized Segmentation using attention-based projection heads.
Experiments and Results
Evaluation on NPC-170 and LIDC-IDRI datasets.
Performance metrics including GED, Dicesoft, Dicemax, Dicematch, DiceA(i), and Dicemean.
Discussions
Visual results showcasing diversified and personalized segmentations.
Selection of hyperparameters K and β for improved performance.
Conclusion
Summary of the study's contributions in addressing multi-rater medical image segmentation challenges.
Estatísticas
Existing works aim to merge annotations or generate diverse/pesonalized results (NPC-170).
Extensive experiments demonstrated superior performance for multi-rater medical image segmentation (LIDC-IDRI).