Multi-focus image fusion is an information fusion technique that combines multiple images of the same scene, each captured with different focal depths, into a single all-in-focus image. This class of methods holds significant practical value in domains such as medical imaging, remote sensing, and photography. To address the limitations of existing multi-focus image fusion methods in terms of feature preservation, feature complementarity, and multi-scale information extraction, this paper proposes a novel fusion approach based on Denoising Diffusion Probabilistic Models (DDPM), named MCGFusion. The proposed method introduces two innovative modules. First, a Parallel Multi-Scale Attention Module is designed to perform multi-scale feature decomposition and enhancement on individual images. By incorporating two parallel convolutional branches operating at different scales, the module effectively captures diverse image features. A simple yet effective pixel-wise attention mechanism is then employed to enhance these extracted features. Next, the decomposed features from the two source images are fed into a Cross Gated Fusion Module, where feature interactions across the two images enable feature refinement. A gating mechanism is subsequently used to adaptively fuse these optimized features, allowing the noise predictor within the DDPM framework to better exploit the semantic content and thereby improve fusion quality. Experimental results demonstrate that MCGFusion achieves superior performance across multiple key metrics, particularly excelling in detail enhancement, mutual information retention, and structural consistency.