Brain Tumor Radiogenomic Classification using multilevel contrastive information fusion

Brain tumors present significant challenges in diagnosis and treatment due to their complex and diverse nature. Radiogenomic analysis, which combines genomic and radiological data, has become important for advancing brain tumor research, with deep learning methods, particularly fusion-based approaches, making notable progress. However, current approaches face challenges in capturing both sequential and spatial features, as well as selecting effective fusion techniques. The research introduces a multi-fusion framework that combines Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and transformers using contrastive learning to enhance feature representation. This approach processes medical images through a neural network alongside a transformer backbone, merging outputs at various stages to improve dark knowledge differentiation between classes. While CNNs extract spatial features, transformers focus on patch-specific information, collectively boosting the model’s discriminative power. The method demonstrates a 2-3% performance improvement over state-of-the-art models, with significant advancements in glioblastoma diagnosis and treatment. An ablation study further confirms that each component positively contributes to the classification task.