Copyright © 2026 Authors retain the copyright of this article. This article is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

@article{206709,
        author = {B Madhavi and Malla Mokshitha Sri Priya and Mamidi Vani Vineesha and Gudivada Avinash and konchada Ashok Kumar and Addanki Pavan Sri Harsha},
        title = {-GRIT-Lite: A Probabilistic Reasoning Framework for Multi-Pathology Kidney Disease Diagnosis from 3D Computed Tomography},
        journal = {International Journal of Innovative Research in Technology},
        year = {2026},
        volume = {13},
        number = {2},
        pages = {2876-2888},
        issn = {2349-6002},
        url = {https://ijirt.org/article?manuscript=206709},
        abstract = {Accurate diagnosis of kidney disease from three-dimensional computed tomography (CT) remains challenging because clinically relevant findings are often small, co-occurring, and anatomically dependent. Existing deep-learning pipelines typically optimize segmentation or classification in isolation, which limits multi-pathology localization, weakens relational reasoning between anatomical structures, and yields poorly calibrated confidence estimates that hinder clinical adoption. We pro-pose C3-GRIT-Lite (Contextual, Causal, Cross-Scale Generative Reasoning Transformer Lite), a unified end-to-end architecture that integrates volumetric self-supervised representation learning, set-based 3D instance detection, probabilistic anatomical graph reasoning, and consistency-constrained multi-task optimization in a single framework. The core novelty is a fully probabilistic graph edge formulation that jointly incorporates learned feature affinity, soft anatomical atlas priors, uncertainty suppression, and detection confidence weighting, preventing the propagation of unreliable detections into downstream relational reasoning. An explicit inter-module consistency loss further enforces prediction coherence across the detection, graph-reasoning, and classification heads. Evaluated on KiTS23, C4KC-KiTS, and CT-KIDNEY, C3-GRIT-Lite achieves a mean Dice score of 0.852, [email protected] of 0.683, macro-averaged AUROC of 0.931, and Expected Cal-ibration Error of 0.031, outperforming all baseline methods across every reported metric. These results demonstrate that explicit anatomical structure, probabilistic message passing, and coherence constraints are essential ingredients for trustworthy 3D renal imaging AI.},
        keywords = {kidney disease diagnosis, 3D computed tomog-raphy, probabilistic graph neural network, uncertainty quantifi-cation, self-supervised learning, multi-pathology detection, con-sistency constraint, DETR},
        month = {July},
        }