Osteosarcoma Classification using Multilevel Feature Fusion and Ensembles

Abstract

Osteosarcoma is a type of bone cancer found in adolescents. Identifying the type of tumour from the histopathological images is a difficult task for the pathologist. In this work, a deep learning based osteosarcoma classification algorithm using ensemble approach and fusion approach is proposed. Multilevel features are extracted from a pre-trained EfficientNets trained on imagenet1k dataset. EfficientNets are scaled convolutional neural networks. This scaling is done in depth, resolution and width. Features are extracted from the initial layers, intermediate layers and final layers of a selected EfficientNet. In general, they represent the low frequency, middle and high frequency details of the images. Independently, the features are given to an error control output coding classifier with support vector machine as base learner. Ensemble prediction is done on the test images by using majority voting from the models trained using features extracted at various levels from EfficientNet. Further, a fused feature vector is formulated from the selected layers of EfficientNets and given to the error control coding output classifier. The proposed algorithm with ensemble approach and fusion approach offers higher mean and peak classification accuracy compared to the existing works in the literature.