Childhood Medulloblastoma Classification Using EfficientNets

Abstract

In this work, a deep learning methodology for accurate classification of histological images of the patients suffering from childhood medulloblastoma is proposed. Pre trained EfficientNets trained on the ImageNet dataset are considered in this work. Features are extracted from the average pooling layer of EfficienNets and are given to an error correcting output code classifier. Ensemble prediction from the selected pre-trained EfficientNets is employed. For the multi class classification, the proposed approach is able to predict with a mean classification accuracy of 98.78% for 10x level images and 95.67% for the 100x level images for an 80% train and 20% test split. The peak classification accuracy is 100% for both binary and multiclass case at cell level and architectural level. For the binary classification with same split, 100% mean classification accuracy is achieved even without ensemble prediction. The results are compared with an existing work on a similar dataset and the significant improvement is demonstrated with the experimental simulations.