The segmentation of the prostate gland into two sub-regions, namely, the central gland (CG) and the peripheral zone (PZ) is crucial for the prostate cancer (PCa) diagnosis. The nature and occurrence of cancer occurred in the prostate is substantially different in both zones. Magnetic resonance imaging modality (MRI) is a clinically primary tool for computer-based assessment and remediation of various cancer types such as PCa. In this paper, we evaluated DeeplabV3+ model on T2W MRI scans using the I2CVB dataset, which is designed in an encoder-decoder style for the zonal segmentation of prostate regions. An important feature of DeeplabV3+ is the depth-wise separable convolutions, which allow more information to be extracted from images as it uses filters with different dilation rates. Prior to being fed to the deep neural network, image pre-processing techniques are applied, including image resizing, cropping, and denoising. The DeeplabV3+ model performance is evaluated using the Dice similarity coefficient (DSC) metric and compared with the vanilla U-Net architecture. Results show that the encoder-decoder network having depth-wise separable convolutions performed better prostate segmentation than the network with standard convolution operations with the DSC value of 70.1 in PZ and 81.5 in CG zone. © 2021, Springer Nature Switzerland AG.