Deep Transfer Learning-Based Multi-class Classification of Bone Tumors Using Medical Imaging

Identifying different types of bone tumors from Xray images is still a tough task for doctors, mainly because many tumors look alike and the visual clues can be extremely subtle. To help with this challenge, our study explores a deep learning approach that classifies bone tumors into four groups: Normal, Normal Tumor, Benign Tumor, and Malignant Tumor. We gathered a secondary X-ray dataset and balanced it using a simple median-based method that mixes undersampling with SMOTE oversampling, and then split it into training, validation, and testing sets. We experimented with four well-known pretrained models—MobileNetV3, VGG19, ResNet50, and EfficientNetB3—each fitted with the same custom classification head that used global average pooling, a ReLU layer, and a Softmax output for the final predictions. In each experiment, EfficientNetB3 did remarkably well. It achieved a test accuracy of 91.13% and showed more consistency when handling tumor classes that are often confused with one another. The ROC curves and confusion matrices provided additional support for its stability. Comparing our results with previous studies, we found that our approach maintained a reasonable computational cost while producing noticeable improvements. Overall, the results indicate that an EfficientNetB3-based model can be a useful and trustworthy tool for helping medical professionals diagnose and classify bone tumors in detail.