Cross-Domain Robust Pruning for Polyp Segmentation: Multi-Encoder Feature Fusion Beats Single-Encoder Baselines

Area of Science:

• Computer Vision
• Medical Imaging
• Machine Learning

Background:

• Medical image segmentation demands extensive pixel-level annotations, increasing dataset construction costs.
• Data-efficient training methods are crucial for developing robust segmentation models.

Purpose of the Study:

• To evaluate if fusing complementary pretrained image encoders enhances training-free dataset pruning robustness.
• To quantify the robustness of this approach across heterogeneous polyp segmentation domains.

Main Methods:

• Propose Multi-Encoder Diverse Pruning (MEDP), a training-free method fusing ResNet-18 and DINOv2 ViT-S/14 features.
• Utilize Louvain modularity maximization for training pool partitioning and Maximal-Marginal Relevance (MMR) for sample selection.
• Benchmark MEDP against 12 baselines at a 20% retention ratio across Kvasir-SEG, CVC-ClinicDB, and a combined cross-domain pool.

Main Results:

• MEDP achieved the highest mean test Dice score of 0.7324 on the challenging Combined cross-domain pool.
• MEDP significantly outperformed uniform random sampling (Cohen's d = +1.79, p = 0.002).
• Hand-crafted structure-aware methods failed to outperform uniform random sampling.

Conclusions:

• Combining multi-encoder features with MMR diversity offers a robust and effective strategy for medical imaging segmentation.
• The choice of pretrained image encoder is a critical factor in segmentation-aware pruning.