FMCL: a transformer-based feature-map classifier learning approach for enhanced brain tumor detection in MRI

In order to achieve a higher level of precision in the diagnosis of brain tumors using magnetic resonance imaging (MRI) data, both traditional image processing and advanced machine learning techniques are utilized. This enables adaptive identification of regions based on relationships between extracted features. In this article, we propose a Feature-Map Contrast Learning (FMCL) framework for improved brain tumor identification. Unlike conventional CNN- or transformer-based methods, FMCL explicitly enforces an inverse proportional relationship between contrast and intensity during feature extraction, preventing feature domination in heterogeneous or low-contrast tumor regions. A proportionate correlation between contrast and intensity is maintained across a large number of classification cases. FMCL performs a multi-class classification of brain MRI images, distinguishing multiple tumor types and healthy tissue by integrating transformer-based self-attention over regulated feature maps, which enhances spatial coherence and tumor boundary localization. During classification, transformer learning differentiates high, low, and zero classification possibilities, and adjacent-region verification with SoftMax normalization further improves segmentation reliability. By utilizing maximal alterations across feature maps, classed regions are accurately recognized and segmented as tumor-infected. To achieve optimal detection precision, FMCL is trained using multiple classification outputs while traversing pixels. The proposed FMCL approach demonstrates improved performance, enhancing accuracy by 8.85% and 8.3%, while reducing mean error by 16.19%, compared to existing methods. The proposed FMCL framework was evaluated using the Kaggle Brain Tumor MRI Dataset, which contains 3264 MRI images of 512 × 512 pixels, including classes for No Tumor, Glioma, Meningioma, and Pituitary tumors.