MUMBAI, India, June 30 -- Intellectual Property India has published a patent application (202641075846 A) filed by Vellore Institute Of Technology on June 19, 2026, for A Selective Feature-Layer And Soft-Label Fused Knowledge Distillation System For Lightweight Medical Image Classification.
Inventors include Arumuga Arun R; Priyanka N; Sabarinathan P; and Manushi Swapnil Bombaywala.
The application for the patent was published on June 26, 2026, under issue no. 26/2026.
Abstract: A Selective Feature-Layer and Soft-Label Fused Knowledge Distillation system (100) for lightweight medical image classification is disclosed. The system (100) comprises a preprocessing module configured to process medical image data, a teacher neural network configured to generate hierarchical feature representations and soft-label probability distributions, a student neural network comprising separable convolution layers, a feature representation alignment module, and a training module. The feature representation alignment module performs selective transfer of intermediate and deep semantic feature representations from non-consecutive teacher layers to corresponding student layers using Global Average Pooling, trainable projection mapping, and L2 normalization. The training module optimizes the student neural network through a multi-objective loss function comprising classification loss, knowledge distillation loss, and feature-alignment loss. The system enables generation of a lightweight classification model through selective feature-layer alignment and soft-label fusion-based knowledge transfer.
Disclaimer: Curated by HT Syndication.
