Enhancing Brain Tumor Diagnosis: Transitioning From Convolutional Neural Network to Involutional Neural Network

Accurate classification of brain tumors is essential for effective medical diagnosis and treatment planning. Traditional approaches rely on convolutional neural networks (CNNs) for tumor detection, but they often suffer from high computational demands due to the large number of parameters. In this paper, we propose a novel approach for brain tumor classification using involutional neural networks (InvNets), which are designed to mitigate the parameter-intensive nature of CNNs. Unlike the spatial-agnostic and channel-specific convolution kernel, the involution kernel is location-specific and channel-agnostic. This location-specific operation allows the network to adapt to various visual patterns with respect to different spatial locations, enhancing its ability to capture intricate features within the medical images. Our study focuses on a four-class brain tumor classification problem, aiming to differentiate between different tumor types based on medical imaging data. In a comparative analysis, we demonstrate that conventional CNNs require over 4 million parameters, whereas our proposed InvNets require less than 0.2 million parameters, making them more efficient and resource-friendly. The evaluation of both CNNs and InvNets is carried out using standard performance matrices: accuracy, precision, recall, F1 score, and AUC-ROC values. Our findings reveal that the InvNets consistently outperform traditional CNNs. The InvNet architecture achieves an impressive 92% accuracy rate, showcasing its potential for accurate brain tumor classification. This improved accuracy, combined with the reduced parameter count, highlights the effectiveness of InvNets for medical image analysis tasks, especially in scenarios with limited computational resources.