Direct Volume Rendering is a popular technique for visualization of 3D datasets that offers many advantages over other algorithms but requires important computing power. Efficient parallelization is therefore an essential need. Taking advantage of an existing LAN instead of using a dedicated parallel architecture can be a cost-effective solution. However this requires special attention in algorithm design to the specific aspects of a LAN as a parallel computing system (long latency, slow throughput). Besides, due to their large volume (up to 1GB), 3D datasets as those produced by modern medical scanners can not be made entirely resident into main memory. We present in this paper the study of an algorithm for direct volume rendering of large images on a local network of workstations which does not assume the presence of the whole dataset in main memory. A method for handling large distributed 3D image datasets is proposed and three parallelization strategies are compared based on communication performance measurements of a LAN.