Brain tumor classification based on SFFM-ResNet18

Brain tumors are becoming more common, making it crucial to accurately classify tumor images for timely diagnosis. To address the limitations of existing deep learning models, such as insufficient feature extraction capabilities, susceptibility to overfitting, and the potential neglect of feature correlations, this study proposes the SFFM-ResNet18 model. Based on ResNet18, the proposed framework integrates a Spatial-Freq- Fusion-Mixer (SFFM) that incorporates spatial-domain, frequency-domain, and Single-Head Self-Attention (SHSA). The model captures both small details and broader patterns, helping it focus on what truly matters for accurate classification, thereby enhancing its global modeling capability. Experimental results demonstrate that the SFFM-ResNet18 outperforms mainstream models in terms of accuracy (99.57%), precision (99.56%), recall (99.56%) and F1-score (99.56%). Visualization analysis further validates its ability to accurately extract critical features with strong robustness. In conclusion, the SFFM-ResNet18 model provides an efficient and reliable solution for brain tumor image classification, offering promising applications in medical image analysis and contributing to advancements in brain tumor diagnosis and treatment.