Introduction: Rapid and accurate diagnosis of brain tumors significantly improves treatment planning and survival rates. However, the manual review of multi‑modal magnetic resonance images is often slow and prone to errors due to lesion heterogeneity, similarity to healthy tissue, and large data volumes. This study presents an automated framework that incorporates a lightweight feed‑forward neural network with an intrinsic attention mechanism and quantum‑behaved particle swarm optimization (QPSO). The aim of this study is to improve the speed and accuracy of tumor identification while maintaining interpretability in clinical environments with limited resources.
Method: Multimodal MRI images (T1, T1ce, T2, and FLAIR) were sourced from reputable databases, including the Brain Tumor Segmentation Challenge (BraTS) and The Cancer Imaging Archive (TCIA). The images underwent preprocessing, which included intensity normalization (Z-score), noise reduction using Gaussian and median filters, and correction of intensity inhomogeneity. Statistical, textural, and frequency-based features were extracted and reduced to 300 principal components using Principal Component Analysis (PCA). Feature weighting was performed using a document relevance-inspired method. The proposed model, a five-layer feedforward neural network (FNN) with a ReLU activation function and an internal attention mechanism, was optimized using QPSO. Heatmaps were generated to enhance result interpretability.
Results: The proposed model achieved an accuracy of 99.6 %, sensitivity of 99.4 %, and specificity of 99.7 %, outperforming reference convolutional networks (97.1 %) and U‑Net architectures (96.2 %). The mean prediction time per image was less than 0.5 seconds, facilitating real‑time clinical use. Heatmaps produced by the attention layer, effectively highlighted abnormal regions and enhanced interpretability. These metrics were consistently replicated across multiple random splits, and qualitative evaluations by imaging specialists confirmed the absence of specificity loss and the clinical relevance of the findings.
Conclusion: A feedforward network equipped with intrinsic attention and optimized with QPSO demonstrated near‑perfect accuracy and sub‑second inference for brain tumor diagnosis on multi‑modal MRI. Its high performance on standard GPUs, combined with the generation of intuitive heatmaps, positions this framework as a practical decision‑support tool, particularly in centers lacking advanced infrastructure. Future evaluations will focus on multi‑center data and deployment on edge devices to strengthen clinical adoption and regulatory compliance.
Rights and permissions | |
![]() |
This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License. |