Text

Objectives and questions: Accurate classification of acetabular fractures is essential for surgical planning but remains challenging due to the complexity of fracture patterns and the subjective nature of interpretation. The Letournel-Judet classification system is the clinical gold standard; however, it is characterized by low inter-observer reliability and a steep learning curve, particularly for less experienced surgeons. These factors may lead to inconsistent diagnoses and treatment decisions. Artificial intelligence (AI) offers the potential to assist in the objective classification of such complex injuries. This preliminary study examines inter-rater agreement between a senior surgeon, a junior surgeon, and a 3D deep learning model in the classification of three acetabular fracture types based on computed tomography (CT) data.

Material and methods: A total of 259 CT scans were retrospectively collected and classified according to the Letournel-Judet system by a senior surgeon, serving as the reference standard for training and evaluation. From the ten original fracture types, a subset of the three most frequent classes was selected (n = 203) as they represented the majority of cases: anterior column with posterior hemitransverse (ACPHT, n = 96), associated both columns (ABC, n = 76), and anterior column (AC, n = 31). The AI model consisted of a 3D convolutional neural network pre-trained on a multi-domain, multi-organ, and pathology dataset. It was first fine-tuned on a pelvic fracture dataset for three-class classification (no fracture, acetabular fracture, other fracture) and subsequently fine-tuned to classify the three selected Letournel-Judet fracture types. Model performance was evaluated on a held-out test set using accuracy and macro-averaged F1 score to account for class imbalance. Pairwise inter-rater agreement among all raters was assessed using Cohen’s kappa (κ).

Results: The three-class Letournel-Judet fracture model achieved an overall accuracy of 0.73 (macro F1 = 0.64). Per-class F1 scores were 0.75 for ACPHT, 0.78 for ABC, and 0.40 for AC. Inter-rater analysis revealed substantial agreement between the junior and senior surgeons (κ = 0.70). Agreement between the AI model and the senior surgeon was moderate (κ = 0.53), as was agreement between the AI model and the junior surgeon (κ = 0.46).

Discussion and conclusions: Despite limitations related to the relatively small dataset and the restricted subset of fracture classes, the AI model demonstrates encouraging performance. However, its agreement with senior expert assessment does not yet match that of the junior surgeon. Analysis of per-class F1 scores indicates that the model primarily struggles with underrepresented classes, particularly anterior column fractures. These findings highlight the potential of AI to support acetabular fracture classification and motivate future studies with larger datasets and an expanded set of fracture classes to enhance reliability and clinical applicability.