<p>Diabetic retinopathy (DR) is a complication of diabetes that affects the blood vessels in the retina. Over time, high blood sugar levels can damage these vessels, causing them to leak or close off. This can lead to vision loss or blindness if left untreated. Segmentation is a crucial task for improving DR diagnosis, but because the lesions are small, asymmetrical, and blurry in morphological patterns, it is complicated, and automated segmentation is challenging. Majority of multi-step segmentation techniques now in use are based on spatial-based categorization with huge computation time, incorporate errors at various stages, and can result in error aggregation. We introduce a computerized segmentation that uses convolutional neural network (CNN) and vision transformer, capture deeper 3D voxel characteristics, and merge the multi-level metadata of scan images. Two components make-up the suggested 3D edge network (3DECNN): (1) The vision transformer modified with a deep multi-resolution attention mechanism (DMRA) is employed as the foundation system; this is the initial effort to channel the lesions using a 3D CNN. (2) A heavy state space transformer architecture is incorporated further into DMRA that emphasize the regional levels and parallelly suppress the unnecessary additional context information. The contextual information in-between the saliency maps at high levels and the voxels at various scales can be dynamically re-aligned. Using data involving 397 individuals with 3200 images taken from the RFMiD database, we conducted tests using five cross-fold validations. The findings prove that the proposed paradigm is effective, more substantial, robust, and more accurate than previous approaches, with the dice coefficient and the sensitivity score being 0.9176 and 0.9846, respectively.</p>