A Lightweight Attention-Based Deep Learning Framework for Brain Tumor Classification Using MRI Images

Abstract

Brain tumors are a serious medical concern that necessitate prompt and precise categorization in order to effectively treat them. Reliable automated solutions are needed since traditional MRI analysis is labor-intensive and might be unpredictable. A lightweight approach combining multi-stage information fusion and an attention mechanism is suggested to accomplish accurate subtype classification of brain tumors while utilizing low resources. Nevertheless, there has been little investigation into how certain attention mechanisms, such as selective kernel Attention Mode (SKAM), affect brain tumor categorization. This research intends to fill that need by classifying brain tumors using mobilenetv3 in conjunction with SKAM (mobilenetv3-SKAM). A multi-stage fusion technique, with mobilenetv3 serving as the backbone, is employed to extract the wealth of information present in images of brain tumors. To further improve the extraction of both underlying and semantic characteristics, an improved squeeze-and-excitation coordinate attention (SCA) mechanism is united in the latter stages of feature extraction, and the selective-kernel (SK) attention mechanism is introduced in the beginning stages. At each step, attention map weights are calculated, and these determine the final feature representations used for subtype categorization. Impressive performance indicators such as 99.93% accuracy, 99.51% recall, 99.23% precision, 99.27% F1-score, and 99.34% AUC were achieved by the proposed framework, outperforming standard deep learning classification approaches on the same dataset. The suggested architecture enables trustworthy automated decision-making in neuro-oncology and offers a scalable and effective method for the real-time clinical diagnosis of brain tumors.