Early　brain　tumor　detection　and　diagnosis　are　critical　to　clinics.　Thus　segmentation　of　focused　tumor　area　needs　to　be　accurate,　efficient,　and　robust.　In　this　paper,　we　propose　an　automatic　brain　tumor　segmentation　method　based　on　Convolutional　Neural　Networks　(CNNs).　Traditional　CNNs　focus　only　on　local　features　and　ignore　global　region　features,　which　are　both　important　for　pixel　classification　and　recognition.　Besides,　brain　tumor　can　appear　in　any　place　of　the　brain　and　be　any　size　and　shape　in　patients.　We　design　a　three-stream　framework　named　as　multiscale　CNNs　which　could　automatically　detect　the　optimum　top-three　scales　of　the　image　sizes　and　combine　information　from　different　scales　of　the　regions　around　that　pixel.　Datasets　provided　by　Multimodal　Brain　Tumor　Image　Segmentation　Benchmark　(BRATS)　organized　by　MICCAI　2013　are　utilized　for　both　training　and　testing.　The　designed　multiscale　CNNs　framework　also　combines　multimodal　features　from　T1,　T1-enhanced,　T2,　and　FLAIR　MRI　images.　By　comparison　with　traditional　CNNs　and　the　best　two　methods　in　BRATS　2012　and　2013,　our　framework　shows　advances　in　brain　tumor　segmentation　accuracy　and　robustness.