Change detection is an important task to rapidly identify modified areas, in particular when multi-temporal data are concerned. In landscapes with complex geometry such as urban environment, vertical information turn out to be a very useful knowledge not only to highlight changes but also to classify them into different categories. In this paper, we focus on change segmentation directly using raw 3D point clouds (PCs), to avoid any loss of information due to rasterization processes. While deep learning has recently proved its effectiveness for this particular task by encoding the information through Siamese networks, we investigate here the idea of also using change information in early steps of deep networks. To do this, we first propose to provide the Siamese KPConv State-of-The-Art (SoTA) network with hand-crafted features and especially a change-related one. This improves the mean of Intersection over Union (IoU) over classes of change by 4.70%. Considering that the major improvement was obtained thanks to the change-related feature, we propose three new architectures to address 3D PCs change segmentation: OneConvFusion, Triplet KPConv, and Encoder Fusion SiamKPConv. All the three networks take into account change information in early steps and outperform SoTA methods. In particular, the last network, entitled Encoder Fusion SiamKPConv, overtakes SoTA with more than 5% of mean of IoU over classes of change emphasizing the value of having the network focus on change information for change detection task.