Transformers in surgical artificial intelligence: A domain-stratified, study-level narrative review

Area of Science:

• Artificial Intelligence in Medicine
• Surgical Technology
• Machine Learning Applications

Background:

• Transformer architectures are increasingly explored in various medical domains.
• Evaluating the comparative performance of transformers against non-transformer models and human benchmarks in surgery is crucial.

Purpose of the Study:

• To synthesize outcomes of transformer-based AI in surgical settings.
• To identify contexts where transformers outperform non-transformer models and human benchmarks.

Main Methods:

• A systematic search of major scientific databases was conducted.
• Eligible studies focused on transformer architectures in surgical/perioperative contexts.
• A narrative synthesis was performed, with within-study deltas calculated for direct comparisons.

Main Results:

• Transformers outperformed non-transformer models in 75% of medical imaging tasks and 82.4% of workflow recognition tasks.
• Advantages in prognosis and education were promising but heterogeneously reported.
• Performance gains were typically single-digit percentage points in comparable settings.

Conclusions:

• Transformer AI frequently matches or exceeds non-transformer baselines in surgical imaging and workflow tasks.
• Further research requires standardized benchmarks, external validation, and studies addressing real-world deployment constraints.
• Ensuring fairness across patient and learner groups is essential for clinical and educational impact.