Optical Coherence Tomography Angiography (OCTA) is a recently developed noninvasive imaging technique capable of capturing detailed images from distinct layers of the retinal vascular complexes. This imaging modality extracts images from the capillary level, consisting of thin, tiny, and high-complexity vascular information. Therefore, segmentation of the retinal microvasculature in OCTA images is challenging. To accurately segment the vessels in OCTA images, we have applied an improved version of RIMNet. RIMNet is a convolutional neural network (CNN) based image magnification architecture. In this study, we have used a depthwise separable convolutional layer instead of the standard convolutional layer in the RIMNet model dubbed as d-RIMNet and experimented on the Optical Coherence Tomography Angiography Retinal Scans and Segmentations (OCTA-SS) dataset. We have noticed that using depthwise separable convolutional layers in d-RIMNet reduces the parameter and complexity of the model by \({\approx }20\%\) than RIMNet without significantly decreasing the segmentation performance. A thorough performance analysis of the state-of-the-art models and our model on the OCTA-SS dataset demonstrates that with a small dataset, smaller network sometimes outperforms large complex networks.