MUMBAI, India, Feb. 13 -- Intellectual Property India has published a patent application (202641008190 A) filed by Koneru Lakshmaiah Education Foundation; Kosaraju Madhavi; Ramakrishna Akella; Prasanna Lakshmi Akella; and Veera Venkata Raghunath Indugu, Hyderabad, Telangana, on Jan. 28, for 'system and method for assessing diabetic retinopathy using lesion-proxy guided weakly supervised framework.'
Inventor(s) include Kosaraju Madhavi; Ramakrishna Akella; Prasanna Lakshmi Akella; and Veera Venkata Raghunath Indugu.
The application for the patent was published on Feb. 13, under issue no. 07/2026.
According to the abstract released by the Intellectual Property India: "Diabetic retinopathy (DR) is a leading cause of preventable vision loss, making accurate severity grading essential for effective screening and timely intervention. Although recent deep learning approaches have achieved promising performance in automated DR grading, most methods either rely solely on image-level supervision with limited interpretability or require costly pixel-level lesion annotations. Moreover, commonly used attention mechanisms do not explicitly capture clinically meaningful lesion cues, restricting their reliability in clinical settings. This paper presents a lesion-proxy guided weakly supervised framework for interpretable DR severity grading using fundus images. The proposed method introduces a lesion-proxy learning module that encodes surrogate lesion-related representations directly from image-level DR grades, without requiring explicit lesion annotations. An EfficientNet-B3 backbone is employed for robust feature extraction, followed by the lesion-proxy module that generates soft activation maps highlighting regions contributing to disease severity. The framework is trained using a multi-objective loss that jointly optimizes grading accuracy and lesion-proxy consistency, enabling the model to learn lesion-aware representations under weak supervision. Experiments are conducted on the large-scale EyePACS dataset for DR severity grading, with additional validation on the IDRiD dataset to assess the clinical relevance of the learned lesion-proxy maps. The proposed framework demonstrates improved grading performance over baseline EfficientNet models and conventional attention-based approaches, while providing more interpretable visual explanations aligned with pathological lesion regions. The results indicate that lesion-proxy guided learning offers an effective and annotation-efficient solution for automated DR severity grading, supporting its potential applicability in large-scale clinical screening environments."
Disclaimer: Curated by HT Syndication.
