Abstract
The segmentation accuracy of deep learning-based brain tumor MRI images still requires further improvement. We proposed a conditional diffusion network that incorporates image information into the mask's perturbed diffusion process. By optimizing the introduction of conditional supervision signals and employing an attention mechanism, our model accelerated convergence and improved predictive performance on the BraTS 2020 dataset. In the public MRI brain tumor segmentation dataset, both performance metrics have improved, with Dice metric increasing by approximately 1.99% compared to the second best metric and IoU metric increasing by 1.61% compared to the second best metric. This suggests the model may provide more stable MRI segmentation, potentially supporting clinical decision-making in research settings.