Content-Based Image Retrieval (CBIR) systems are widely used for local as well as for remote applications such as telemedicine, satellite image transmission and image search engines. The existing CBIR systems suffer from the limitations of storage space, data security and bandwidth requirement. To overcome these problems, a new method termed as CBIR-PE which makes use of prediction errors instead of actual images for storage, transmission and retrieval is presented. Identical artificial neural networks (ANNs) are employed both at the server and client sides to carry out the prediction. At the server side, only the error database comprising the difference between the original and the predicted pixel values is used instead of the actual image database. The prediction errors of the query image are matched with those in the server database to retrieve similar prediction error patterns. These errors are then combined with the predicted values available at the client ANN to reconstruct the actual images. Since only the prediction errors are employed, the proposed method is able to solve the problems of storage space, data security and bandwidth requirement. The proposed method is implemented in combination with a clustering technique called WBCT-FCM which makes use of wavelet based contourlet transform (WBCT) and fuzzy c-means (FCM) clustering algorithm. The performances of the proposed WBCT-FCM and CBIR-PE are evaluated using COREL- 1k database. The experimental results show that the proposed methods achieve better results with respect to clustering and retrieval accuracies compared to the existing methods.