SEIMB-NET: A Squeeze and Excitation Driven Lightweight Model for Classification of Brain Tumors Using Magnetic Resonance Imaging

Brain tumor is one of the prevailing, degenerative, and progressive brain disease for adults and children all over the world. The conventional method of evaluating brain tumor progression through manual examination of magnetic resonance images (MRI) by a radiologist may result in errors, impacting the chances of survival of the patient. Due to the different characteristics of brain tumors, it is difficult to classify the different types of tumors. Early tumor type identification often relies on automated end-to-end frameworks, frequently employing deep neural networks to detect tumors using MR images. However, to tackle the scarcity of MR images for training Deep Learning models, transfer learning techniques utilize pre-trained models such as ResNet, VGGNet, DenseNet, and InceptionNet for classifying tumors using brain MRI scans. But deploying transfer learning models in resource-constrained medical devices is challenging due to their large size. Therefore, to address this problem various pre-trained lightweight models are proposed, though working with these models may occasionally compromise the overall network performance. To address this issue, we designed a lightweight deep neural network utilizing a fine-tuned MobileNetV1 with an optimized architecture that employs depth-wise separable convolution which reduces the size of the proposed model, and integrated it with a Squeeze and excitation (SE) attention block for effective brain tumor classification. The proposed work outperformed different state of the art models in existing literature utilizing different evaluation measures, achieving an accuracy of 98.56%, precision of 98.37%, recall of 98.36%, and F1 score of 98.38% with an optimum model size of 14MB.