Author ORCID Identifier
0009-0005-6472-3173
Biosketch
Dr.Renuga Devi M is a researcher in Artificial Intelligence and Medical Image Analysis with expertise in Deep Learning, Reinforcement Learning, Radiomics, and Brain Tumor MRI analysis. Her research focuses on glioma segmentation, survival prediction, explainable AI, and healthcare analytics using advanced AI techniques. She has published research articles in reputed journals and publishers including MDPI Diagnosis, Wiley, and IEEE. She is also actively involved in interdisciplinary research, teaching, and AI-driven healthcare innovation.
Date of Award
30-3-2026
Document Type
Thesis
School
School of Electrical & Electroncis Engineering
Programme
Ph.D.-Doctoral of Philosophy
First Advisor
Dr.K.Narasimhan
Keywords
Reinforcement Learning, Deep Learning, Medical Imaging, Glioma Prognosis, Brain Tumor Survival Prediction
Abstract
Brain tumors are highly aggressive and lethal types of cancer, particularly gliomas. These cancerous growths show complicated pathophysiological behaviours and with poor prognosis despite therapeutic advances. Due to their biological differences and infiltrating growth, as well as overlapping radiological characteristics, they pose great difficulty in diagnosis, grading, and survival prediction. Artificial intelligence technology, which includes machine learning , deep learning, and reinforcement learning has developed into a new paradigm for the automation of brain tumor diagnostics and personalized treatment. The main goal of this study is to create an integrated AI-based framework that can perform brain tumor segmentation, grading and survival prediction via hybrid ML, DL and RL paradigms. Various challenges in the literature, such as data scarcity and imbalance, high-dimensional heterogeneity, low interpretability, and lack of non-adaptive and sequential learning methods for continuous survival estimation motivate this work. To perform continuous survival prediction using reinforcement learning, leveraging radiomics, deep features, and explainable AI for patient-specific prognosis.
This study aims to fill these gaps by designing explainable and adaptive AI approaches driven by multi-modal MRI datasets and dynamic learning mechanisms to enhance clinical outcome prediction. The research unfolds in four major technical phases. In the first phase, a hybrid framework based on Machine Learning-Deep Learning was developed for tumor segmentation, classification, and survival prediction using BraTS2020. An enhanced UNet++ network was used to segment the tumor sub region, achieving a testing accuracy of 98% and an IoU score of 0.7483. Data balancing (SMOTE, ADASYN) and dimensionality reductions (PCA, tree-based selection) were used on the radiomics features obtained from the segmented tumor region. Out of the classifiers, the Stochastic Gradient Descent (SGD) model achieved 96.87% accuracy, while XGBoost regression recorded a Mean Square Error of 93726.45 for survival prediction.
The second phase introduced hybrid vision U-net (HVU) architectures, HVUED for segmentation and HVU-E for classification, by integrating u-net with ViT and pretrained CNN backbone ResNet50, VGG16, DenseNet121 and Xception. These architectures managed to capture local context and global context through multi-scale feature extractor. The DenseVU-ED model obtained Dice scores of 0.902, 0.954 and 0.966 for the enhancing tumor, core and whole tumor regions, respectively, while DenseVU-E’s classification accuracy was 99.18%. Explainable AI (XAI) techniques like Grad-CAM, LIME, and SHAP improved interpretability by highlighting clinically relevant tumor sub regions and features. In the third phase, a framework GlioSurvQNet based on reinforcement learning was proposed for tumor classification and discrete survival prediction under fully and data-limited scenarios.
A metaheuristic ensemble of the Harris Hawks Optimization, the Modified Gorilla Troops Optimization, and the Zebra Optimization were employed to optimize the multimodal MRI-based radiomics features along with SHAP for interpretability. The best modality FLAIR + T1CE accurately classified LGG–HGG at 99.27%, and FLAIR + T2 + T1CE predicted survival at 93.82% accuracy. The DuelContextAttn DQN model showed remarkable stability and robustness under data-scarce conditions. With a grading accuracy of 100% and a survival prediction accuracy of 94.47%, the Brier scores suggest they are well-calibrated and reliable.
The last step was to design a SimCLR-Twin Critic Deep Deterministic Policy Gradient (DDPG) framework for continuous survival prediction that moves from discrete to continuous outcome model. A self-supervised reinforcement learning system that combines contrasting representation learning (SimCLR) with a twincritic DDPG agent capable of adaptive policy learning from continuous feedback. The highest segmentation accuracy was obtained by nnU-Net among the features extracted from U-Net, V-Net, and SwinUNETR. The combination of radiomics and deep features was refined using BorutaShap for dimension reduction. The experimental outcomes showed an improvement over DeepSurv, regular DDPG, and standard machine learning models. It has a C-index of 0.87, MSE of 31062.36 and MAE of 96.73. LIME-based explanations suggest that tumor shape, texture, and patient age affect survival predictions.
In concisely, this thesis proposes an AI framework that is comprehensive, interpretable, data-efficient, and capable of adaptive learning for personalized prognosis modeling. The self-supervised and reinforcement learning paradigms provide novel insights to improve medical imaging analytics and move beyond static classification toward continuous survival estimation. The models under study demonstrate potential for integration into clinical workflows as decision-support systems, where MRI scans can be automatically analyzed for tumor segmentation, grading, and survival prediction. The explainable AI components further assist clinicians by highlighting important tumor regions and influential features, thereby supporting diagnosis, treatment planning, and survival management in neuro-oncology.
Recommended Citation
M, Renuga Devi Ms, "An Interpretable Hybrid Deep and Reinforcement Learning Paradigm for Glioma Prognosis" (2026). Theses and Dissertations. 204.
https://knowledgeconnect.sastra.edu/theses/204