Diabetic Retinopathy (DR) is a serious complication of diabetes mellitus that involves progressive deterioration of the blood vessels in the retina and can result in permanent vision loss. The stealthy development of the disease requires vigorous early diagnosis mechanisms. This paper presents MSRA-Net, a new Convolutional Neural Network (CNN) architecture that is specifically designed to diagnose diabetic retinopathy from fundus images automatically. This methodology involves the utilization of Mish activation functions (M) for stability in training, Spatial Attention modules (S) to prioritize pathologically relevant areas, Residual Blocks (R) to retain important features, and AdaptiveAvgPool2d (A) to retain global spatial information. The APTOS 2019 Blindness Dataset is used and employs preprocessing methods such as Contrast Limited Adaptive Histogram Equalization (CLAHE) and Gaussian blur for image quality improvement. Empirical validation asserts that MSRANet accomplishes 96.1% accuracy, surpassing a number of state-of-the-art deep learning methods. Architectural excellence of the proposed network in feature extraction and classification has encouraging prospects for clinical applicability in early detection frameworks for DR, which can be used to alleviate the public health burden of diabetes-related vision loss.