Aiming at the low accuracy of traditional brain tumor detection, a three-dimensional brain tumor detection method based on deep learning was proposed. Firstly, the magnetic resonance images of different modal brain tumors were fused to obtain the three-dimensional features of brain tumor focus under different modalities. Then, an instance normalization layer was added between the convolution layer and the pooling layer to improve the convergence speed of the network and relieve the problem of overfitting. And the loss function was improved, the weighted loss function was used to enhance the feature learning of the focus area. Finally, the problem of more focuses in the false positive brain tumor was solved combining with the post-processing method. The experimental results show that the proposed brain tumor detection method can effectively detect the tumor focuses. The Dice coefficient, sensitivity and specificity of the three evaluation indexes reach 0.926 7, 0.928 1 and 0.997 7 respectively. The three indicators improve 4.6%, 3.96% and 0.04% compared with the 2D detection network, and improve 13.2%, 10.42% and 0.12% compared with the initial single modal brain.