Optimization of MRI- Based Residual Network for Prediction of Brain Tumor Disorder Using Transfer Learning

Diagnosing abnormal growth of tissues lying within brain at an initial stage has become very crucial task to enhance outcomes of patients in the healthcare. Most of the conventional techniques depend upon expertise of medical professionals for interpretation of achieved outcomes which can lead to time-consuming process along with false predictions especially when tumor size is too small at an initial stage. This research focuses on the optimization of magnetic resonance imaging based extended version of residual network, ResNet50V2 for prediction of brain tumor disorder where residual blocks contain several convolutional layers, basic blocks, bottleneck blocks, shortcut connections responsible for minimizing computational cost and providing direct pathway for gradients to flow while training network following transfer learning approach to leverage parameters learned from ImageNet objects. In addition, support vector machines have been integrated with residual network for performing classification among healthy and infected images. Afterwards, validation is performed on the proposed model based on test data and gained 98.92% accuracy, 98.15% precision, 99.69% sensitivity, 98.18% specificity and 98.91% Dice-coefficient. Thus, hyperparameters along with performance-metrics obtained in the proposed model signify the remarkable outcomes as compared to the existing approaches while diagnosing brain tumor disorder in the healthcare.