Performance Analysis of Vision Transformer (ViT), ResNet50, and MobileNetV3 Large in Multiclass Bone Fracture Classification

Ei Phyu Sin Win (Mandalay Technological University, Myanmar)
Phyo Thu Zar Tun (Mandalay Technological University, Myanmar)

The Indonesian Journal of Computer Science
Vol. 15 No. 3 (2026): The Indonesian Journal of Computer Science

Automated classification of bone fractures has become a cornerstone of modern emergency radiology, significantly enhancing diagnostic speed and precision. This study evaluates the comparative efficacy of three leading deep learning frameworks ResNet50, MobileNetV3, and Vision Transformer (ViT) using a diverse dataset that includes various fracture modalities, healthy X-rays, and non-radiological images.The experimental data reveals that the Vision Transformer (ViT) attained the highest diagnostic accuracy at 95%, marginally outperforming MobileNetV3 and ResNet50, which both achieved 94%. While all three models demonstrated flawless reliability (100%) in identifying Forteen Classes Bone categories, their performance diverged when analyzing complex fracture patterns.