MUMBAI, India, July 7 -- Intellectual Property India has published a patent application (202621057780 A) filed by Nishtha Garg on May 06, 2026, for A Calibrated Multimodal Fusion Approach For Early-Stage (stage 0–1) Diabetic Retinopathy Detection.
Inventors include Amruta Bhawarthi; Niyati Parmar; Nishtha Garg; Atharva Padwal; Nirmayee Ninawe; and Aditya Pallerla.
The application for the patent was published on July 03, 2026, under issue no. 27/2026.
Abstract: ABSTRACT The early diagnosis of Diabetic Retinopathy (DR) is very important in clinical practice, but it is difficult to distinguish between Stage 0 (no DR) and Stage 1 (mild DR) based on early signs like microaneurysms, which are sparse and subtle. Most current research on DR focuses on multi-class classification of the severity level, in which Stage 1 always has low sensitivity and poor calibration performance. This paper offers a multimodal fusion approach tailored for binary Stage 0/Stage 1 DR detection by fusing retinal fundus images with clinical biomarkers. The proposed approach consists of three independently trained models: (i) an image- only classifier using EfficientNet-B4, (ii) a tabular model trained with LightGBM on five clinical variables (HbA1c, fasting glucose, cholesterol, diabetes duration, and age), and (iii) a late fusion meta-learner using logistic regression. The outputs are Platt-scaled to address probability mis-calibration before fusion. Uncertainty estimates based on entropy are also introduced into the fusion feature space to enhance robustness and mitigate overconfident predictions. The experimental results show that, although the image-only and tabular-only models have complementary strengths and weaknesses in early-stage detection, the proposed fusion approach leads to better recall for Stage 1 DR, enhanced probability calibration, and more robust decision boundaries. The proposed approach is also validated on an external test set, showing better generalization performance than unimodal baselines. The interpretability of the proposed approach is provided by Grad-CAM visualization for retinal image classification and SHAP analysis for clinical biomarker contributions. This paper exclusively targets Stage 0/Stage 1 DR discrimination and thus emphasizes the power of multimodal fusion and calibrated decision-making for one of the most challenging and less explored problems in DR detection using automated computer vision approaches.
Disclaimer: Curated by HT Syndication.
