JMIR Publications

ABSTRACT

Background:

Neuroimaging segmentation is crucial for the diagnosis and treatment planning of a variety of neurological conditions. Manual segmentation is time-consuming and subject to human error and variability. Transformers are a promising deep-learning approach for automated medical image segmentation.

Objective:

This scoping review aimed to synthesize current literature and assess the use of different transformer models for neuroimaging segmentation.

Methods:

A systematic search was conducted across four major databases (Scopus, IEEE Xplore, PubMed, and ACM Digital Library) for studies applying transformers to neuroimaging segmentation problems. Relevant articles were screened for inclusion based on our inclusion criteria, and data extraction was performed to identify key study details, including image modalities, datasets used, neurological conditions transformer models, and evaluation metrics used.

Results:

Among the 1246 publications identified, 67 (5.38%) met our inclusion criteria and were therefore included in this review. Half of all included studies were published in 2022, with over two-thirds of them applying transformers for brain tumor segmentation. The most commonly used imaging modality and dataset were Magnetic Resonance Imaging (MRI) with almost 90%, and BraTS with 58%. Three-dimensional transformer models were favored over two-dimensional variants, with the hybrid CNN-transformer architecture being the most extensively developed.

Conclusions:

This review demonstrates the recent growth in implementing transformers for neuroimaging segmentation. Hybrid CNN-transformer approaches currently achieve the best results on benchmark datasets compared to standalone models. Further research into the field and the development of more standardized datasets are required to further advance the application of transformers in medical imaging. Clinical Trial: Brain tumor segmentation, 3D segmentation, Transformer, Deep learning