With the development of artificial intelligence, numerous computer-aided diagnosis systems (CADSs) have been proposed to diagnosis meningiomas and gliomas automatically. However, most current systems not only ignore the large intra-class variances among original brain images, but also employ small databases with expensive labeling costs; as a result, the performances of most CADSs are well below expectations. To optimize the diagnosis performance of meningiomas and gliomas, novel image processing methods, including a novel multi-directional brain region extraction (MDBRE) method and an iterative gamma correction based on two peaks (TPGC), are proposed to narrow the intra-class variances, and a pre-trained AdaBound-SE-DenseNet (AD-SE-DenseNet) is presented to avoid over-fitting. First, innovative image processing methods, including a novel MDBER and a novel TPGC, are applied to remove the disturbances of skulls and brightness variances. Then, data augmentation technologies are applied to produce a larger database and a pretrained AD-SE-DenseNet is introduced to train the classifier. The experimental results indicate that the accuracy of this system can reach 96.87%. Implementing the innovative image processing methods and AD-SE-DenseNet can lead to a nearly 8% and 1.7% accuracy improvement, respectively.