Deformable Convolution Network for Reconstruction of Misaligned MRI Images and Classification of Brain Tumor in the Aligned Images

Magnetic Resonance Imaging (MRI) is an popular imaging approach which is increasing its employment for preclinical research of many diseases. Brain Tumor is considered as one of largest deadliest cancer diseases across the world. MRI provides excellent soft tissue information in deep imaging range towards efficient detection and classification of the regions of interest. Especially MRI images were widely employed in brain tumor detection applications. However MRI images leads to several challenges due to misaligned images and image distortion. Hence it becomes mandatory to reconstruct the image quantitatively and qualitatively with strong generalization. In this article, a deformable convolution network is employed for detection and reconstruction of the misaligned MRI images on incorporation of the deformable convolution and deformable pooling layer to analyze the geometric variations of the MRI images and further it is used for detection and classification of the brain tumor diseases. Deformable Convolution Network (DCN) comprises of restoration network which accurately aligns the MRI images using reinforcing method and self attentive network which extract rich features on reinforced image regions for detection and classification of the brain tumor diseases. Deformable Convolution Network is capable of encapsulating the structure of the various elements to attain the accurate and reliable feature extraction and classification. Proposed DCN model enables offset learning with deformable convolutions uses 2D offset to sampling locations to identify the deformation on the preceding feature map. Next, Deformable max pooling enables adaptive part localization for tumor regions. Experimental analysis of MRI image reconstruction and image classification approach is carried out using Stanford 2D dataset and BRATs dataset in the Matlab environment. In this analysis, proposed deep learning model considered to be excellent compared to traditional classifier models. Finally performance evaluation of the current deep learning based image reconstruction approach and existing MRI image reconstruction approach is carried out using Alignment accuracy , Sensitivity and Specificity metrics, it reveals that proposed reconstruction methods achieves better performance compared to conventional models on specified metrics.