MUMBAI, India, Aug. 24 -- Intellectual Property India has published a patent application (202641096430 A) filed by Rajalakshmi Engineering College on August 10, 2026, for Hog Assisted Transformer And Convolutional Neural Network Fusion Architecture With Encoder Decoder Segmentation For Automated Skin Lesion Classification.
Inventor includes Dr. K. Anand.
The application for the patent was published on August 21, 2026, under issue no. 34/2026.
Abstract: The detection of melanoma at an early stage significantly increases patient survival, yet complex lesion structures and inter-class variability present primary challenges in interpretation. This invention introduces a novel HOG-assisted Transformer-CNN fusion network with U-Net segmentation for robust melanoma classification. The model initiates with a U-Net—based segmentation layer to isolate diagnostically relevant lesion regions, followed by the integration of three parallel encoders: a Convolutional Neural Network (CNN) to capture deep spatial features, a Histogram of Oriented Gradients (HOG) encoder to identify fine-grained texture patterns, and a lightweight Vision Transformer (ViT) to model long-range global dependencies. A feature fusion module integrates these local and global representations, enhancing discriminability between similar lesion types. Experimental validation on ISIC 2016, 2017, and 2018 datasets demonstrates that the proposed architecture outperforms traditional CNN, ResNet50, and vision-transformer baselines. The method achieves significant gains in precision, AUC, and Fl-score while maintaining performance under variable illumination and class imbalance. Furthermore, Grad-CAM visualization is integrated to highlight lesion regions consistent with dermatological ABCDE criteria, providing the explainability required for clinical adoption. The findings confirm the model's suitability for accurate melanoma classification and high-precision computer-aided diagnosis in real-world clinical settings.
Disclaimer: Curated by HT Syndication.
