<p>In the field of Reversible Data Hiding (RDH), effective management of image prediction and prediction errors is critical. This paper proposes a novel approach that integrates a Convolutional Neural Network (CNN)-based predictor for precise pixel value estimation with a two-phase embedding process. The method generates a difference image by computing prediction errors and segments it into uniformly sized blocks. In the first phase, the extreme right difference value predicts left difference values, enabling modifications on the left side based on secret data bits. In the second phase, the extreme left prediction difference value forecasts right difference values, allowing adjustments on the right side according to secret data bits. These modified prediction difference values are then incorporated into the original image to produce the stego-image. Experimental results demonstrate that the CNN predictor achieves superior prediction accuracy compared to existing predictors. Additionally, the proposed RDH method outperforms conventional RDH techniques in embedding performance.</p>