Brain tumors represent a significant challenge to public health worldwide. The early detection and diagnosis are crucial for improving patient life expectancy. The unique methodology of this study consists of exploring and comparing two different approaches: “black box” deep learning and hand-crafted feature-based machine learning, showing how artificial intelligence can be helpful in clinical workflows in different ways. The proposed methods achieve classification accuracies of 99% with XceptionNet, ensuring at least a 3% improvement from existing approaches, and 94% with Soft Voting, respectively. The discussion shows that the deep learning approach is more accurate than its machine learning counterpart but less interpretable. The practical applicability of the proposed approach is highly evaluable because both the transfer learning and machine learning models employ low computational costs and can be effectively implemented even in resource-constrained clinical environments.