Multiscale Deep Feature Extraction for Brain Tumor Detection using Hybrid CNN and Wavelet Transform

In medical imaging, spotting brain tumors is crucial since an accurate and precise diagnosis has an enormous impact on the outcome of treatment and prognosis for patients. This work proposes an innovative approach to improve the accuracy of brain tumor recognition by merging convolutional neural networks, or CNNs, with a deep wavelet autoencoder (DWA). Model generality across various tumor types is ensured by integrating a strong data augmentation mechanism to handle data paucity. The CNN algorithm identifies malignant and non-tumor areas with high precision, while the DWA efficiently eliminates noise and extracts multiscale features. When compared to conventional techniques, experimental results show greater performance, with notable gains in specificity, sensitivity, and absolute classification accuracy. Brain tumor identification can be enhanced by combining architectures such as VGG, ResNet, and DenseNet with Discrete Wavelet Transform (DWT). The proposed framework achieved: Accuracy: 97.5%, Sensitivity: 96.8%, Specificity: 98.3%. DWT enhances the hierarchical characteristics that CNNs extract by emphasizing the combination of structural (low-frequency) and comprehensive (high-frequency) information. By integrating the benefits of several designs, fusion increases overall robustness and lessens dependency on a single model. Merging several deep learning models raises the amount of computing complexity, which requires careful manipulation of hyperparameters, overcome by the attention mechanism.