Concetti Chiave
Proposing a polyp segmentation model based on Prompt-Mamba with Vision-Mamba technology and prompt assistance for improved generalization capabilities.
Sintesi
Introduction:
Importance of polyp segmentation in medical image processing.
Challenges in accurately segmenting polyps due to various shapes, colors, and unclear boundaries.
Existing Models:
Deep learning models like U-Net, U-Net++, ResUNet, PraNet, TGANet, SAM for polyp segmentation.
Limitations of existing models in handling irregular shapes or small samples.
Vision-Mamba Technology:
Migration of Mamba architecture from NLP to CV domain.
Performance of Vision-Mamba and its application in medical image segmentation.
ProMamba Model:
Structure with image-encoder, prompt-encoder, and mask-decoder.
Utilization of box-prompt for prompting tasks.
Experimental Results:
Training on merged datasets and testing on unseen datasets.
Comparison with SOTA methods showing superior performance by 5% across datasets.
Ablation Study:
Impact of varying vision-Mamba layers and embedding length on model performance.
Ablation experiments showcasing the importance of backward_SSM and input_mask components.
Conclusion:
Proposal of a novel segmentation model combining Vision-Mamba and prompt technologies for enhanced generalization capabilities in polyp segmentation.
Statistiche
"Compared to previous models trained on the same dataset, our model not only maintains high segmentation accuracy on the validation part of the same dataset but also demonstrates superior accuracy on unseen datasets."
"Our code and trained weights will be released soon."
"Our results are 5% better than the performance of the previous best methods."
Citazioni
"We propose a concise and effective model structure that incorporates prompt technology, resulting in outstanding generalization capabilities."