A Hierarchical Transformer with Residual CGTM Blocks and Gated Dilated Convolutional Refinement for Binary Bone Fracture Classification from Radiographic Images

Abstract

Automated bone fracture classification from radiographic images remains challenging because fracture evidence is often subtle, spatially confined, and sensitive to projection angle and anatomical variation. While convolutional models and standard transformer-based architectures have advanced radiographic image analysis, they may not jointly capture fine-grained local structural disruptions and the broader contextual dependencies required for reliable fracture recognition. In particular, existing methods do not sufficiently unify localized fracture-sensitive feature extraction, cross-window contextual interaction, and multi-scale structural refinement within a single classification framework. To address this gap, this study proposes a hierarchical transformer-based network for binary bone fracture classification built on residual CGTM blocks. The architecture integrates window-based multi-head self-attention, shifted-window multi-head self-attention, and a Gated Dilated ConvFFN within a four-stage representation hierarchy, with adaptive patch merging between stages and attention pooling for discriminative feature aggregation. The model was evaluated on a bone fracture radiograph dataset using dedicated training, validation, and held-out test subsets, with performance stability assessed through 10-fold cross-validation. Comparative evaluation was conducted against MaxViT and EVA-02 under identical experimental conditions. The proposed model achieved a mean validation accuracy of 99.276% and a held-out test accuracy of 99.25%, together with a ROC-AUC of 99.99% and an average precision of 99.99%, outperforming MaxViT at 97.99% test accuracy and EVA-02 at 93.355% mean validation accuracy. These results indicate that jointly modeling local structural sensitivity and cross-window contextual interaction within a hierarchical transformer framework offers a meaningful advantage for radiographic fracture classification, though broader validation remains necessary before clinical use.