Strategies for Deep Learning in Volumetric Medical Imaging: A Survey

• João Vitor S. de Oliveira UFV
• Danilo F. Vieira UFV
• Mateus P. da Silva UFV
• Daniel L. Fernandes UFV
• Marcos H. F. Ribeiro UFV
• Hugo N. Oliveira UFV

Resumo

Deep learning has transformed medical image analysis. However, building effective models for volumetric data, such as CT, MRI, and PET scans, presents a new set of challenges. These include high computational costs, limited annotated datasets, and the need for architectures that can process volumetric data directly. This survey provides a comprehensive overview of recent advances in deep neural network architectures specifically designed for 3D medical imaging. We analyze the progression of architectural choices along the years, highlighting their innovations and applicability. In addition, we review training strategies such as supervised/self-supervised pretraining and the development of general-purpose 3D foundational models. Also, a comprehensive overview of 3D medical imaging datasets and their associated tasks is presented in the supplementary materials, highlighting the diverse clinical objectives that deep learning models can support across healthcare and diagnostic applications. A dedicated section addresses the emerging field of explainability in volumetric contexts, emphasizing the limitations of adapting 2D explainability-focused tools and the importance of 3D-native explanation frameworks. We conclude by outlining the key trends in the field of 3D medical imaging.

Palavras-chave: Deep learning, Surveys, Training, Solid modeling, Technological innovation, Three-dimensional displays, Reviews, Computer architecture, Self-supervised learning, Biomedical imaging, 3D medical imaging, Explainable AI, Volumetric segmentation