Deep learning and hybrid architectures for atypical and complex bone fracture diagnosis: a systematic review of performance and clinical validity.

Artificial intelligence (AI) is reshaping fracture diagnosis in medical imaging. Despite these advances, accurately identifying atypical fractures (such as stress or pathological fractures) and complex fractures (including comminuted and pelvic fractures) remains a significant clinical challenge. This systematic review evaluates the current evidence on AI models, including advanced architectures, for detecting, classifying, and segmenting atypical and complex bone fractures in humans. A total of 40 studies published between 2015 and 2026 met the predefined inclusion criteria. Eligible studies used real-world imaging modalities (X-ray, CT, or MRI), focused on atypical or complex fractures, employed AI-based approaches with expert-validated reference standards, and reported quantitative performance metrics. Studies based exclusively on synthetic data, restricted to simple fractures, or lacking adequate validation were excluded. Advanced AI models, including hybrid frameworks such as 3D U-Net variants and DeepLabV3+MobileNetV3, were associated with improved performance in several studies, particularly for identifying subtle and multi-fragment fractures. However, substantial heterogeneity in study design, datasets, validation strategies, and evaluation metrics limits direct comparisons across models. Hybrid systems, particularly CNN-based architectures combined with level-set methods or multi-network pipelines, also appeared effective in capturing complex fracture patterns in several studies, although this observation is based on a limited and heterogeneous body of evidence. Overall, the available evidence suggests that advanced AI models have considerable potential to improve the detection, classification, and segmentation of atypical and complex fractures. Nevertheless, the predominance of single-center studies, the limited use of external or prospective validation, and methodological heterogeneity indicate that further standardized, multicenter clinical validation is required before these models can be widely implemented in routine clinical practice.