Microscopic medical image segmentation plays a vital role in computational pathology by enabling the analysis of cellular and subcellular structures. However, the presence of noise, low contrast, and overlapping regions makes the task highly complex. This paper proposes a novel segmentation framework that combines a diffusion-based generative model with a U-Net architecture, enhanced by Dempster-Shafer Theory (DST)-based feature fusion. The diffusion component models the progressive addition and removal of noise, training the network to reconstruct clean anatomical structures across varying noise levels. The DST-based fusion module is incorporated at each decoder stage to refine feature integration under uncertainty, allowing the model to selectively focus on reliable spatial regions. This design enhances the model’s robustness and structural consistency, particularly in noisy or ambiguous regions. The proposed method is evaluated on three benchmark datasets – MoNuSeg, TNBC, and Kumar demonstrating significant improvements over existing techniques. Dice Scores of 84.36%, 88.33%, and 84.43% are achieved on MoNuSeg, TNBC, and Kumar datasets respectively, surpassing several state-of-the-art models. These results highlight the effectiveness of combining probabilistic diffusion modeling with uncertainty-aware fusion for accurate and generalizable segmentation in complex medical imaging scenarios. The code implementation of the methodology is available at: DDU-Net .