Intellectual Property India Publishes Patent Application for An Attention-Augmented Hybrid Convolutional Neural Network System And Apparatus For Real- Time Explainable Brain Tumor Diagnosis And Pathological Tissue Localization

MUMBAI, India, Sept. 22 -- Intellectual Property India has published a patent application (202641108646 A) filed by Sr University on September 10, 2026, for An Attention-Augmented Hybrid Convolutional Neural Network System And Apparatus For Real- Time Explainable Brain Tumor Diagnosis And Pathological Tissue Localization.

Inventors include Dr. B Padmaja; and Dr. Maram Balajee.

The application for the patent was published on September 18, 2026, under issue no. 38/2026.

Abstract: AN ATTENTION-AUGMENTED HYBRID CONVOLUTIONAL NEURAL NETWORK SYSTEM AND APPARATUS FOR REAL-TIME EXPLAINABLE BRAIN TUMOR DIAGNOSIS AND PATHOLOGICAL TISSUE LOCALIZATION The present invention discloses an attention-augmented hybrid convolutional neural network system and apparatus (Atten-CNNBrainNet / AttentiveBrainNet) for real-time explainable brain tumor diagnosis and pathological tissue localization from Magnetic Resonance Imaging (MRI) scans. The system comprises a Multi-Stage Preprocessing Engine that resizes scans to 224 x 224 pixels, converts slices to grayscale, normalizes pixel intensities, and applies adaptive inverse class weighting to eliminate dataset class imbalance bias. A Hierarchical Feature Extraction Backbone containing three 2D convolutional blocks (32, 64, 128 filters) extracts multi-scale spatial feature maps. A Custom Spatial Self-Attention Module Computes Query, Key, and Value projections to generate a 2D spatial attention weight matrix, selectively amplifying discriminative tumor region features while suppressing background signals. A Regularized Dense Classification Head with 50% dropout outputs normalized Softmax class probabilities across target diagnostic categories (Glioma, Meningioma, Pituitary Adenoma, Healthy). An Explainable Spatial Heatmap Generation Engine transforms the attention weight matrix into a color-coded JET heatmap overlaid directly onto the raw MRI slice, delivering 98.3% classification accuracy, 96.8% sensitivity, 0.982 ROC-AUC, and 100% auditable spatial explainability within a lightweight parameter footprint (~3.2 million parameters).

Disclaimer: Curated by HT Syndication.