Prostate segmentation is a key step for several clinical applications, both preoperatively and intraoperatively. Efforts to automate it would contribute to a more consistent level of quality for this procedure and reduce inter- and intra-operator variability, thus ensuring better patient care. We introduce 3D-MXF an automatic segmentation approach using ensemble deep learning to merge the segmentation result proposed by several expert networks. Three CNN are trained to specifically segment 2D images extracted from a single 3D TRUS volume according to the three main views (axial, coronal, sagittal). The main contribution lies in the specific fusion step performed by a 3D CNN trained to provide 3D confidence maps that enable the fusion of the volumes reconstructed from the expert networks’ outputs into a final segmented volume. The 3D-MXF framework was trained and carefully evaluated on a database containing prostate TRUS images of 382 patients. The results are superior to what is obtained with other state-ofthe- art methods. A fine analysis of clinical parameters demonstrated the potential and limitations of this approach for a clinical use.