MoDiff is a morphology-emphasized diffusion model designed for ambiguous medical image segmentation. It replaces traditional one-hot encoding with probability-based label maps to capture inherent uncertainties and ensure consistent segmentation results. By determining the presence of individual radiologist labels, MoDiff enables diverse sampling that provides richer insights into ambiguous areas. Its Learnable Discrete Frequency Filter (LDF) extracts high-frequency details for improved boundary precision, and when integrated with the Morphology-based Cross Attention Network (MCA), it enhances feature synthesis for more accurate anatomical segmentation. Evaluations on the LIDC-IDRI and MS-MRI datasets confirm its superior accuracy, boundary precision, and consistency.