MUMBAI, India, Aug. 10 -- Intellectual Property India has published a patent application (202641090656 A) filed by Andhra University on July 25, 2026, for Anterior-Posterior Asymmetry Encoding Scheme For Deep Learning-Based Keratoconus Classification.
Inventors include Mr. Prabhu Teja Geddada; and Prof. P. Rajesh Kumar.
The application for the patent was published on July 31, 2026, under issue no. 31/2026.
Abstract: ABSTRACT OF THE INVENTION: Abstract: Corneal tomography captures keratoconus pathology through multiple anterior and posterior corneal maps,yet existing multiview fusion approaches treat these as unstructured peer inputs, modelling inter-view dependencies via attention or concatenation without explicitly encoding the relational structure between anatomically paired anterior and posterior surfaces. In this an explicit anterior-posterior asymmetry encoding scheme is proposed, in which paired anterior and posterior feature vectors extracted per map type via a shared convolutional backbone are decomposed into two complementary descriptors: a difference vector capturing the direction and magnitude of inter-surface asymmetry, and a sum vector encoding overall bilateral severity. For each of three paired map types (elevation, sagittal curvature, and eccentricity), these descriptors are concatenated alongside an unpaired anterior corneal thickness feature, producing a structured 1792-dimensional representation that directly reflects the clinically established anterior-posterior asymmetry signature of keratoconus. This descriptor is passed through a three-layer MLP for three-class classification of keratoconus, suspect keratoconus, and normal corneas.Evaluated on 423 eyes, the proposed method achieves 88.89% overall accuracy, macro-average AUC of 0.961, and suspect keratoconus recall of 77.8% on a held-out test set of 63 eyes, compared to 82.54% accuracy and 44.4% suspect recall for an unstructured flat concatenation baseline under identical conditions.
Disclaimer: Curated by HT Syndication.
