Deep Learning-Based Automated Computed Tomography Classification of Posterior Malleolar Fractures for a Standardized...

Despite advances in computed tomography (CT)-based classification systems for posterior malleolar fractures (PMFs), diagnostic consistency remains limited because these systems are still highly dependent on the observer's experience. This study aimed to develop a deep learning-based automated model capable of classifying PMFs according to the Haraguchi, Mason-Molloy, and Bartoníček-Rammelt clasification systems using CT scans, and to compare the classification performance of different convolutional neural network architectures. Retrospectively collected preoperative CT data from 480 adult patients with PMFs were analyzed. A total of 2398 CT slices were independently classified by three orthopedic surgeons according to the three classification systems, and majority consensus was used as the reference standard. Anonymized DICOM images were evaluated using VGG-16, ResNet-101, and DenseNet-121 architectures. The dataset was divided into 60% training, 20% validation, and 20% testing subsets. Model performance was assessed using accuracy, sensitivity, precision, F1 score, and area under the receiver operating characteristic curve (AUC). The VGG-16 and ResNet-101 models demonstrated high classification performance in the Haraguchi and Mason-Molloy systems. In the VGG-16 model, AUC values exceeded 0.90 across most fracture subtypes, reaching particularly high levels in Haraguchi types I-III, Mason-Molloy types IIB and III, and Bartoníček-Rammelt type IV. Similarly, the ResNet-101 model achieved AUC values above 0.90 in most classes, with very high discriminative performance observed in selected subtypes. F1-scores were greater than 0.70 among most fracture categories and exceeded 0.80 in several, indicating strong and consistent classification performance. These findings demonstrate that deep learning-based automated CT classification of PMFs is feasible and diagnostically robust. The proposed framework may improve diagnostic consistency, reduce observer-dependent variability, and support standardized assessment of posterior malleolar fractures in clinical practice.