An Enhanced Hybrid CNN–Transformer-Based Clinical Decision Support System for Automated Brain Abscess Detection and Visualization

Abstract

Early and correct identification is needed to treat bacterial brain abscesses, which cause localized inflammation and pus collection in brain tissue. Manual MRI interpretation is time-consuming, expertise-dependent, and subject to inter-observer variability, making automated diagnostic assistance systems crucial. This paper proposes a Hybrid Deep Learning Framework to automatically detect and visualize bacterial brain abscesses in MRI images. Advanced convolutional neural network designs DenseNet121, ResNet50, and InceptionV3 are used to utilize complementary feature learning capabilities such fine-grained texture extraction and multi-scale spatial pattern identification. The system normalizes, resizes, and geometrically transforms images to increase model resilience and generalization across MRI circumstances. Combining backbone network predictions using a hybrid ensemble and fusion technique reduces model bias and improves classification reliability. Grad-CAM++-based visualization generates attention heatmaps that show infected regions influencing model decisions to improve interpretability and clinical confidence. Testing shows that the hybrid model beats CNN architectures in accuracy, precision, recall, and F1-score, obtaining above 95% accuracy on the test dataset. By clearly highlighting suspicious abscess locations, visualization outputs aid AI-assisted diagnosis. This automated system speeds diagnostics and helps radiologists make solid clinical decisions. The suggested method improves AI-driven medical imaging systems for brain abscess identification, risk assessment, and treatment planning