Classification of Childhood Medulloblastoma and its subtypes using Transfer Learning features - A Comparative Study of Deep Convolutional Neural Networks

Convolutional Neural Network (CNN) has exposed great results for classification as well as segmentation problems, for various image datasets. We widen the use of Deep Learning features using Transfer learning for our medical image dataset of childhood medulloblastoma (CMB) tissue samples. Due to the critical nature of CMB studying its characteristics is of profound significance for integration into the digital healthcare system. The experiments were carried out with data collected from collaborative medical centres of the region. This paper performs classification of CMB samples for two categories: binary (to classify it from normal samples) and multiclass (to classify its different subtypes). For feature extraction, two transfer learning networks: Alexnet and VGG16, were trained and then evaluated and compared. Following this, Softmax function was used as the classifier for both networks. Further, the features extracted from these networks were also compared using traditional machine learning Support Vector Machine (SVM) classifier. Data Augmentation was performed to control overfitting of samples by the network. Performance evaluation showed that Alexnet outperforms VGG-16 network with a softmax classifier but the features extracted from VGG-16 showed superior performance for SVM classification. This suggests that the features extracted by the VGG-16 network are more considerable than Alexnet.