In the area of reversible data hiding (RDH), multiple-histograms modification (MHM) has been widely recognized as one of the most high-performance techniques. With MHM, the correlation between the prediction-error (PE) and the local complexity (LC) can be well exploited for pixel sorting based data embedding, which is very important in MHM-based RDH algorithms for the obtained high image quality and the well-preserved embedding capacity. However, since PE and LC are usually obtained using different algorithms, their correlation may not be strong enough. In this paper, a novel correlation measurement is proposed by exploring the relationship of neighboring PEs to improve the performance of pixel sorting. In the proposed work, we first divide a cover image into non-overlapping cells of two pixels, and segment the pixel cells into two layers for layer-wise data embedding. Then, in the process of prediction, pixels in one layer are estimated from pixels in the other layer using a convolutional neural network (CNN), the loss function of which is designed to minimize not only the PE itself, but also the difference of PEs within the same pixel cell. Finally, both the difference of PEs and the LC are employed for pixel sorting to minimize the embedding distortion introduced by MHM. Extensive experiments demonstrate the effectiveness of the proposed method.