Abstract
Scoliosis affects 2–3% of the population, and manual measurement using the Cobb angle — the gold standard since 1948 — is subject to inter-observer variability and too slow for use in mass screening programs. This paper presents a systematic review of the application of deep neural networks (DNNs) in the detection and analysis of scoliosis through an analysis of the relevant literature published between 2015 and 2024, using the PubMed, IEEE Xplore, Scopus, and Web of Science databases. A range of architectures is considered — from classical convolutional neural networks and encoder–decoder models to Transformer and hybrid CNN–Transformer approaches — together with the key datasets, evaluation metrics, and open problems in the field. Technical progress is clearly visible and measurable: recent literature shows that hybrid CNN–Transformer models achieve a Cobb angle estimation error that approaches the level of disagreement between two experienced radiologists. Nevertheless, most published models have been tested exclusively on data from the institution in which they originated, which raises serious concerns regarding external validity — it remains unclear whether these models would achieve the same performance in other hospitals, with different imaging equipment and different patient populations. This is one of the key open questions in the field and an important direction for future research.