Holographic data storage (HDS) is a promising for next generation archival media. In reproduction, reproduced data of two-dimensional data array are image data and can be demodulated accurately by using convolutional neural networks (CNNs) for each modulation block. In this paper, we consider CNNs which are more effective for suppression of inter-symbol interference to improve the demodulation error rate. We designed and evaluated CNNs into which not only the images of modulation blocks but also the adjacent pixels are input.