1 Professor, Thangal Kunju Musaliar Institute of Technology, Kerala, India. Email: rejir@tkmit.ac.in
2 Senior Lecturer, Laurea University of Applied Sciences, Espoo, Finland. Email: mitha.jose@laurea.fi
3 Associate Professor, St. Joseph's College of Engineering, Chennai, India. Email: lathagct97@gmail.com
4 Technical Project Manager, GITEC-IGIP GmbH, 51063 Cologne, Germany. Email: anith.fingent@gmail.com
5 Assistant Professor, KMCT Institute of Technology, Kerala, India. Email: nesarose@gmail.com
6 Research Scholar, Cochin University of Science and Technology, Kerala, India. Email: lekshmi.r.nair@cusat.ac.in
7 Postdoctoral Research Fellow, University of Southampton, Southampton, United Kingdom. Email: d.saleela@soton.ac.uk
8 Assistant Professor, TKM College of Engineering, Kerala, India. Email: s4shyna@gmail.com
Brain tumour subregion segmentation from multi-modal MRI is clinically important for diagnosis, treatment planning, and disease monitoring, yet accurate delineation remains challenging because the whole tumour, tumour core, and enhancing tumour differ markedly in size, appearance, and boundary complexity. Existing approaches often predict these regions in parallel, which can ignore their natural anatomical hierarchy and lead to inconsistent subregion outputs. To address this, we propose HiEDL-MRI, a hierarchical 3D deep learning framework for brain tumour segmentation that explicitly models the progression from Whole Tumour (WT) to Tumour Core (TC) to Enhancing Tumour (ET). The key idea is to use earlier-stage predictions to guide later, more difficult stages, while also incorporating boundary-aware regularisation and evidential uncertainty estimation to improve structural plausibility and confidence awareness. The framework uses a shared 3D encoder with stage-wise decoder heads, where WT prediction informs TC prediction and TC subsequently informs ET estimation. In addition, boundary-weighted loss encourages sharper tumour contours, and an evidential learning component provides uncertainty quantification for the final ET stage. Using multi-modal MRI volumes resized to a unified 3D target space, the model is designed to produce clinically meaningful and logically consistent tumour subregion maps. The proposed framework achieved Dice scores of 0.9000 for WT, 0.8600 for TC, and 0.8000 for ET on the validation set. Overall, the proposed framework is significant because it combines hierarchical segmentation, boundary refinement, and uncertainty modelling within a single end-to-end pipeline, offering a more interpretable and clinically relevant strategy for brain tumour analysis from MRI.
Keywords: Hierarchical segmentation; Brain tumour MRI; Multi-modal MRI; Boundary-aware learning; Evidential uncertainty estimation
How to cite this article: Reji R, Jose MR, Latha P, Shaji A, Sabu N, Nair LR, Saleela D, Shyna A. Hiedl-MRI: A Hierarchical 3D Deep Learning Framework with Evidential Uncertainty Modelling for Brain Tumour Subregion Segmentation from Multi-Modal MRI. Int J Drug Deliv Technol. 2026;16(5): 298-309. DOI: 10.25258/ijddt.16.5.32
Source of support: Nil.
Conflict of interest: None