基于扩散模型的细胞图像与掩膜联合生成数据扩充方法

Key Points

• This research aims to address the insufficiency of annotated data in cell image segmentation tasks through a novel data augmentation method.
• Developed a joint generation framework using a noise-predicting U-Net and a noise sampler.
• Created four-channel joint images by stacking cell images and masks.
• Implemented a single-stage generation process to minimize errors associated with multi-stage methods.
• Enhanced modeling capabilities using time embedding and multi-head self-attention.
• Optimized the training process with mean squared error loss and cosine annealing learning rate scheduling.
• Demonstrated significant performance improvements for downstream segmentation models.
• Achieved IoU and Dice scores of 62.50% and 75.78% respectively on the CryoNuSeg dataset.
• Showed superior performance compared to traditional augmentation methods like flipping and rotation.