Effective segmentation of intracranial hemorrhage (ICH) is key to early diagnosis and treatment of stroke. Traditional segmentation methods rely on manual intervention, which is inefficient and susceptible to subjective factors, while existing deep learning models have limitations in handling ICH lesions of different shapes and sizes. Denoising Diffusion Probabilistic Models (DDPM) have demonstrated effectiveness across various vision tasks, such as image deblurring, super-resolution, and anomaly detection. Motivated by DDPM’s success, this paper introduces a new network, RDT-Net, built upon DDPM. We combined the model with the Multi-Scale Adaptive Rotated Convolution (MARC) module, which improves segmentation accuracy by dynamically adjusting kernel weights, eliminating redundant features, and integrating multiple features and angle information derived from the feature maps. In addition, RDT-Net ensures the uniqueness of stable generated masks through a dynamic conditional encoder with a deep stack transformer (DST), thereby enhancing segmentation accuracy. Experimental results show that RDT-Net outperforms other models significantly, achieving superior results in terms of the Dice coefficient, Jaccard index, and 95% Hausdorff distance, providing an effective and dependable solution for the complex task of ICH segmentation.