Hybrid Deep Learning Framework for Bone Tumor Classification from X-Ray Images Using EfficientNet, DenseNet, and Ensemble Voting

Bone tumor detection using X-ray images is still an
important clinical problem because of the highly heterogeneous
nature of bone lesions in X-rays. In this paper, we propose a
hybrid deep learning model for binary bone tumor classification
using X-ray images. The bone tumor positive X-ray images and
normal bone X-ray images are used as input data for our
task. Three convolutional neural network models are trained
independently: EfficientNetB4, EfficientNetB3, and DenseNet121
using overfitting and without overfitting avoidance strategy. We
propose two hybrid approaches: a feature-level fusion approach
based on EfficientNetB4 and DenseNet121 using Squeeze-and-
Excitation (SE) channel attention module and a hybrid approach
based on EfficientNetB4 with the use of SVM classifiers with
RBF, linear, and polynomial kernel functions. We apply three
ensemble voting schemes, including soft, hard, and weighted
voting methods, for our three backbone networks. For all ten
approaches considered, the best results are obtained by soft and
hard voting ensembles, with 99.19% test accuracy and AUC-ROC
of 0.9996. Visualization with Grad-CAM technique confirms that
the model highlights the regions of interest in the image.