HSADML: Hyper-Sphere Angular Deep Metric based Learning for Brain Tumor Classification

Brain Tumors are abnormal mass of clustered cells penetrating regions of brain. Their timely identification and classification help doctors to provide appropriate treatment. However, Classifi-cation of Brain Tumors is quite intricate because of high-intra class similarity and low-inter class variability. Due to morphological similarity amongst various MRI-Slices of different classes the challenge deepens more. This all leads to hampering generalizability of classification models. To this end, this paper proposes HSADML, a novel framework which enables deep metric learning (DML) using SphereFace Loss. SphereFace loss embeds the features into a hyperspheric-manifold and then imposes margin on the embeddings to enhance differentiability between the classes. With utilization of SphereFace loss based deep metric learning it is ensured that samples from class clustered together while the different ones are pushed apart. Results reflects the promi-nence in the approach, the proposed framework achieved state-of-the-art 98.69% validation accu-racy using k-NN (k=1) and this is significantly higher than normal SoftMax Loss training which though obtains 98.47% validation accuracy but that too with limited inter-class separability and intra-class closeness. Experimental analysis done over various classifiers and loss function set-tings suggests potential in the approach.