Bone age assessment is a routine task in pediatric radiology. The most common bone age assessment method is the Greulich and Pyle (GP), based on a reference population of Caucasian middle-class children in the USA between 1931 and 1942 [1]. This method compares the patient's left-hand radiograph to the GP atlas depending on chronological age and sex. Other methods exist, such as the Tanner Whitehouse [2], but GP is the most commonly used.

Bone age assessment is a diagnostic tool used in pediatric endocrinology to assess skeletal maturity and to determine a child's growth potential, estimate final height, diagnose endocrine disorders, and monitor treatment response. The interpretation of bone age requires a comprehensive understanding of normal bone development and the impact of various pathologies, such as growth hormone deficiency, hypothyroidism, and precocious puberty on skeletal growth.

Apart from its intended clinical applications, bone age assessment has also been used in legal and forensic contexts to determine a person's chronological age in legal procedures of child labor, sexual assault, prostitution, and young asylum seekers [3,4], which raises some ethical issues. The GP atlas was indeed developed to assess skeletal development with knowledge of the chronological age based on a Caucasian population, but its use cannot be reliably extended to other applications [4,5].

Despite its widespread use, the GP method is subjective and is known for its high inter- and intra-reader variability [4,6,7]. Therefore, accurate and reproducible bone age assessment is crucial, even though this estimation will be repeated when following up most patients in order to evaluate the process of bone maturation.

Automated tools have experienced significant growth in the last few years, especially in medical imaging and for different tasks such as patient triage, medical reporting, image quality improvement, or lesion detection. Artificial intelligence (AI) advancements in musculoskeletal pediatric radiology have mainly focused on automated tools for fracture detection [8], [9], [10], [11], [12] and for bone age assessment on conventional radiographs, with the goal to improve accuracy and reduce variability [13], [14], [15], [16], [17], [18], [19], [20], [21], [22], [23], [24], [25]. Increasing interest about bone age algorithms was noted following the Radiological Society of North America machine learning challenge in 2017 [26]. Now several AI solutions are commercially available with promising results in terms of accuracy and efficacy in different settings and populations.

The purpose of this study was to compare the performance of a new AI software and a senior general radiologist for automatic bone age assessment to a standard of reference established by pediatric radiologists.