An integrated hybrid U-Net and EfficientNetV2-S approach for brain tumor segmentation and classification

Brain tumors involve the uncontrolled proliferation of cells either inside or adjacent to brain tissue, which frequently results in poor patient outcomes. Accurate diagnosis and early identification are vital for effective treatment planning. In this study, we develop a sequential deep learning pipeline for automatic brain tumor segmentation and classification using the publicly accessible Figshare dataset, comprising 3,064 images from 233 patients. We present a double hybrid encoder model based on the U-Net architecture, which combines complementary feature extractors to improve segmentation performance. The model achieves a loss of 0.047, an intersection over union (IoU) of 86.89%, and a Dice score of 95.27%, surpassing the performance of conventional U-Net and U-Net++ architectures. For classification, we utilize a modified EfficientNetV2-S, which is lightweight and achieves a 99% F1 score and 99% accuracy, while being less computationally intensive and faster to train than deeper frameworks such as ResNet50V2. Model performance was evaluated using cross-validation, which included fault detection to improve reliability. We propose that our framework can reliably and efficiently analyze brain tumors and serve as an important component in clinical decision-making in neuro-oncology.