Input Paradigms for 3D MRI-Based Computer Vision: A Systematic Review of Datasets, Tasks, and Evaluation Practices

Jiawei Tian¹, Kyungtae Kang^2,*
1 Department of Computer Science and Engineering, Hanyang University, Ansan, Republic of Korea
2 Department of Artificial Intelligence, Hanyang University, Ansan, Republic of Korea
* Corresponding Author: Kyungtae Kang. Email: email
(This article belongs to the Special Issue: The Collection of the Latest Reviews on Advances and Challenges in AI)

Computer Modeling in Engineering & Sciences https://doi.org/10.32604/cmes.2026.087740

Received 22 June 2026; Accepted 26 August 2026; Published online 10 September 2026

Abstract

Deep learning has become increasingly important in magnetic resonance imaging (MRI)-based computer vision, but its application to three-dimensional MRI is still shaped by a basic methodological decision: how volumetric data are represented before model training. A 3D MRI scan may be processed as independent 2D slices, adjacent-slice 2.5D inputs, full 3D volumes, local patches, regions of interest, or multi-view representations. These choices influence spatial-context modeling, computational cost, annotation requirements, architectural design, and task suitability. This review provides a paradigm-oriented overview of recent 3D MRI-based computer vision studies, focusing on the relationships among datasets, input representations, downstream tasks, and evaluation practices. A PRISMA-style literature search was conducted in the Web of Science Core Collection for studies published from 2021 to 2026. After task-oriented screening and manual eligibility assessment, 178 studies were included. The reviewed literature was organized according to major publicly available dataset families, including BraTS, ADNI, OASIS, IXI, HCP, ATLAS, MSD Brain Tumor, OAI, fastMRI, and MIRIAD; input paradigms, including 2D, 2.5D, full 3D, and multi-view designs; and downstream tasks, including detection or diagnosis, segmentation, reconstruction or restoration, and super-resolution. The review indicates that full 3D modeling remains the dominant strategy in volumetric MRI analysis, particularly for segmentation and reconstruction-related applications. However, 2D, 2.5D, and multi-view strategies remain valuable when data, annotations, computational resources, or deployment conditions are limited. Future research should move toward task-adaptive representation, dynamic view selection, lesion-oriented sampling, transparent preprocessing, stronger external validation, and clinically interpretable evaluation. By shifting the focus from individual network architectures to data representation choices, this review provides a reproducible framework for understanding methodological trade-offs and selecting task-appropriate input paradigms for 3D MRI-based computer vision.