metadata
license: apache-2.0
pipeline_tag: image-segmentation
VQ-Seg: Vector-Quantized Token Perturbation for Semi-Supervised Medical Image Segmentation
This model corresponds to the VQ-Seg training setup on the ACDC dataset.
VQ-Seg is the first approach to employ vector quantization (VQ) to discretize the feature space in semi-supervised medical image segmentation. It introduces a controllable Quantized Perturbation Module (QPM) that replaces traditional dropout, enabling effective regularization by shuffling spatial locations of codebook indices.
- Paper: VQ-Seg: Vector-Quantized Token Perturbation for Semi-Supervised Medical Image Segmentation
- Code: GitHub - script-Yang/VQ-Seg
- Dataset (ACDC-PNG): Hugging Face Datasets
Key Features
- Quantized Perturbation Module (QPM): Replaces dropout with a mechanism that shuffles spatial locations of codebook indices for better regularization.
- Dual-branch Architecture: Shares the post-quantization feature space between image reconstruction and segmentation tasks to mitigate information loss.
- Post-VQ Feature Adapter (PFA): Incorporates guidance from a foundation model (DINOv2) to supplement high-level semantic information.
Citation
If you find this work useful in your research, please consider citing:
@inproceedings{yangvq,
title={VQ-Seg: Vector-Quantized Token Perturbation for Semi-Supervised Medical Image Segmentation},
author={Yang, Sicheng and Xing, Zhaohu and Zhu, Lei},
booktitle={The Thirty-ninth Annual Conference on Neural Information Processing Systems}
}