MUMBAI, India, Sept. 22 -- Intellectual Property India has published a patent application (202641109638 A) filed by Vallurupalli Nageswara Rao Vignana Jyothi Institute Of Engineering And Technology on September 12, 2026, for Integrated Hardware-Software System For Offline Dermoscopic Skin Lesion Classification Using Hierarchical Vision Transformer At The Point Of Care.
Inventors include Dr. Srinivas Talasila; Dr. Vijaya Kumar Gurrala; Ch Sri Karthik; B Sathvika; and Dr V Pradeep Kumar.
The application for the patent was published on September 18, 2026, under issue no. 38/2026.
Abstract: An integrated hardware-software system for real-time, edge-deployed, multi-class skin cancer detection is disclosed. The system comprises a Raspberry Pi 5 single- board computer quad core ARM Cortex-A76, 8 GB RAM physically coupled with a 12-megapixel Raspberry Pi Camera Module 3 via a MIPI CSI-2 interface for direct dermoscopic image acquisition at the point of care. At the core of the system is a Swin Transformer Tiny (Swin-T) deep learning model a hierarchical, shifted-window self-attention transformer with 28 million parameters fine-tuned on a curated dermoscopic dataset of 111,708 images spanning 17 clinically significant skin condition classes. The model achieves a test accuracy of 99.73%, macro precision/recall/F1-score of 99.71%, and a perfect AUC-ROC of 1.0000 on the 12,424-image evaluation set, outperforming all CNN baselines (85–91%) and the standard ViT-Base model 98.85%, 86M parameters, unsuitable for edge deployment.To enable real-time inference on the Raspberry Pi 5, the PyTorch-trained model is exported to ONNX format and subjected to INT8 dynamic quantization, reducing model size from approximately 108 MB to 20.66 MB 73.9% reduction with less than 1% accuracy loss. The ONNX Runtime inference engine, configured with full graph optimization and quad-core parallelism, achieves median inference latency of 265 ms and throughput of 3.77 images per second at a stable operating temperature of 59.8 degrees Celsius. A Flask-based local web server hosted on the Raspberry Pi provides a browser-accessible interface enabling dermoscopic image upload, AI inference, and classification result delivery without any internet connectivity. The complete system cost is approximately US$147. The invention is distinguished from prior art in that it employs Swin-T's hierarchical, linear-complexity O(n) shifted-window attention rather than standard ViT's quadratic O(n^2) global attention, which renders ViT fundamentally unsuitable for the edge deployment context due to its 86M parameters, 330 MB footprint, and 4–5 second per-image inference latency on ARM processors.
Disclaimer: Curated by HT Syndication.
