The rapid development of deep learning techniques has led to major advancements in medical image segmentation. The majority of segmentation models now in use are discriminative, i.e., they are mostly aimed at developing a mapping between segmentation masks and the input image. These discriminative techniques, however, suffer from an unstable feature space and ignore the underlying data distribution of input samples. This issue is highly pertinent to the segmentation of microscopic medical images, which often have low contrasts and intricate patterns. This paper suggests at using a generative models understanding of the underlying data distribution to supplement discriminative segmentation techniques. Hence, a diffusion based segmentation model is proposed in this study in combination with the Fast Fourier Transform (FFT). The proposed model integrates diffusion principles in the frequency domain. After that, a U-Net architecture with an FFT-based feature extraction aided by an attention mechanism is designed to enhance the segmentation performance. By combining frequency-domain processing, attention module and iterative noise reduction, the model effectively captures both global and local features, enabling precise segmentation of complex structures in microscopic medical images. The effectiveness of the model has been evaluated on three publicly available standard and complex microscopic medical image datasets. The proposed model has obtained Dice scores of 88.13%, 88.52% and 98.57% on TNBC, CPM17 and GlaS datasets, respectively, which are better than many recently proposed models found in the literature. The code implementation of the methodology is available at: FCAM-Diffusion .