F-DL: Fusion of Deep Learning and Image Patch Segmentation for Brain Tumor Detection

Abstract

The brain's unchecked and fast cell development is what fuels a tumor. It may be deadly if not treated right away. Accurately identifying and classifying it remains difficult despite much effort and encouraging results. It is quite difficult to identify a brain tumor because of the differences in the tumor's location, form, and size. The number of "Brain Tumors (BTs)" is rising quickly worldwide. Deadly brain tumors claim thousands of lives each year. For this reason, proper identification and categorization are crucial to brain tumor therapy. Many methods based on “Deep Learning (DL)” and classical “Machine Learning (ML)” have been developed for BT categorization and detection. It takes a lot of effort to create the hand-crafted features needed by the classic ML classifiers. Conversely, deep learning (DL) has become a popular tool for detection and classification due to its strong feature extraction capabilities. As a result, we presented a model fusion-based convolutional neural network model for brain cancer classification in this research. Enhancement of the model's feature recognition performance was achieved by combining several models, combining deep and shallow features, and adding an attention module. Additionally, several tasks were carried out to improve the model's classification performance, including parameter fine-tuning, data augmentation, and model pre-training. Our fusion model is compatible with ConvNeXt_S and baseline EfficientNet B7 models. The model shows an accuracy of 99.88%.