Deep-TLBO: Achieving Robust Deformable Medical Image Registration Leveraging Deep Learning and Teaching Learning-Based Optimization

Abstract

For a doctor to diagnose a patient and conduct quantitative analysis, medical images must be accurately registered. Deep learning-based image registration methods have been explored extensively, but accurate registration of medical images still remains a major concern. To address the issue of accurate alignment in medical images, this paper presents an approach employing unsupervised learning algorithm U-shaped convolution neural network (U-Net) model, followed by teaching learning-based optimization (TLBO) along with affine transformation for grayscale as well as colored medical image registration. The combined two-channel image generated using moving and fixed image is given as an input to encoder and decoder phase of U-Net model to learn image features and generate displacement field. To predict the spatial transformation parameters, a set of control points are generated to define the deformation field and moving image feature to produce the warped image. To improve the quality of warped image, TLBO with rigid transformation parameter (RTP) on fixed and U-Net warped image is applied by detecting the optimum value of transformation parameters. The proposed approach is implemented and evaluated using five different datasets for 2D and 3D monomodal medical MR and CT image modalities as well as multimodal clinical datasets. For 3D representation, transfer learning is used to obtain the warped images using the 3D pre-trained weights of VoxelMorph U-Net model. In comparison to state-of-the-art approaches like symmetric image normalization (SyN) and VoxelMorph, the Dice score value increases from 0.742 as reported by Balakrishnan (Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2018) to 0.9542 and 0.710 as reported by Zhou (IEEE 18th International Symposium on Biomedical Imaging (ISBI), 2021) to 0.9024, while the Hausdorff distance decreases from 2.0521 Zhang (Appl Intell 12:1-18) to 1.5607 and 2.0446 Zhang (Appl Intell 12:1-18) to 1.0923, respectively, for deformable medical image registration under different image modalities. Higher value of Dice score and lower value of Hausdorff distance with our proposed approach in similarity metrics indicates better registration accuracy.