2D, 2.5D, or 3D? Comparing Dimensional Approaches in Deep Neural Networks for 3D Medical Image Analysis.
Authors
Affiliations (3)
Affiliations (3)
- Department of Radiology, Xinhua Hospital Affiliated to Shanghai Jiao Tong University School of Medicine, Shanghai, China.
- Department of Radiology, Songjiang Hospital Affiliated to Shanghai Jiao Tong University School of Medicine, Shanghai, China.
- Department of Radiology, Shanghai TCM - Integrated Hospital, Shanghai University of Traditional Chinese Medicine, 230 Baoding Road, Hongkou District, Shanghai, 200082, China. [email protected].
Abstract
Deep learning has become an important tool in medical image analysis, including 3D modalities like CT, MRI and PET/CT. However, there is no consensus on the most effective dimensionality for neural network models to process this volumetric data. This review analyzes the performance of 2D, 2.5D, and 3D deep neural networks in 3D medical imaging tasks. We synthesized findings from 31 comparative studies published in the last decade, focusing on classification and segmentation tasks. Among them, nearly all studies agreed that pure 2D models perform suboptimally and are generally not preferred. Of the 21 studies that compared 2.5D and 3D architectures, 12 suggested that 2.5D models outperformed other models, 5 favored pure 3D models, and 4 concluded that the relative performance depends on the specific task context. Further analysis reveals that pure 2D models generally underperform due to their failure to capture inter-slice spatial context. While pure 3D networks can theoretically leverage complete volumetric information, their performance is often constrained by practical challenges, including high computational demands and imperfect training data. Conversely, 2.5D models, which balance spatial information with computational efficiency through various strategies, frequently achieve comparable results. This review suggests that no single dimensionality is universally optimal. The choice between a 2.5D and a 3D approach should be guided by the specific clinical task, dataset characteristics, and available computational resources.