Optimizing Brain Tumor Detection: Simplifying U-Net with Low-Rank Factorization, Depthwise Separable Convolution, and MobileNet Architecture

Deep learning models have become essential tools in brain tumor detection and classification. However, high param-eter count of deep learning models regularly leads to overfitting and high computational demands, restricting their deployment in real scenarios. This research presents a detailed analysis to optimize brain tumor detection by reducing model complexity while improving performance. In this study three key concepts are explored: Low-Rank Factorization through Singular Value Decomposition (SVD), Depthwise Separable Convolutions (DSC), and the integration of Efficient Network Architectures such as MobileNet. Low-Rank Factorization efficiently reduces the pa-rameters by approximating the weight matrices, saving important information while decreasing redundancy. Depthwise Separable Convolutions replace standard convolutions with parameter-efficient convolutions, decreasing the model size and computation and at the same time not compromising on accuracy. Finally, adding MobileNet's architecture leads to better model that allows faster and resource-conscious training and interpretation. Exper-iment results shows that these methods drastically avoids risk of overfitting and largely improves accuracy. These developments improve the application of deep learning models in real-time. When this model applied for brain tumor detection accuracy is increased from 97.2% to 97.8^%.