DNN-Reg: A Deep Neural Network for Brain Hemorrhage Deformed CT Image Registration

Medical image registration aligns two or more images and is a crucial step in the automated analysis of medical images. To avoid manual registration methods, we propose a deep learning-based DNN-Reg architecture for brain hemorrhage deformed CT image registration. The proposed architecture includes a CNN-based regressor, a spatial transformer, and a resampler to perform image registration. A sequential CNN consists of convolution, pooling, and fully connected layers, which are used as regressor. This CNN is highly effective in predicting both affine (rigid) and b-spline (non-rigid) spatial transformation parameters and displacement fields. These are used to transform the deformed image close to original image via spatial transformer and resampler (bicubic interpolation). CNN updates its weights through back propagation when the resampler fails to achieve sufficient similarity between two images. The proposed model operates in an unsupervised manner and is trained end-to-end by optimizing the similarity between deformable images and corresponding original images. Once trained, the DNN-Reg model can register unseen image pairs. Using publicly available Kaggle brain hemorrhage CT dataset, the proposed DNN-Reg model of affine and deformable b-spline registration achieves the best test accuracy of 84.7%, mean square error (MSE) of 12.2% and normalized cross-correlation (NCC) of 87%. The proposed model enhances radiology-based image analysis and serves as a valuable tool to support clinicians in accurate diagnosis.