基于差分曲率分组混合模型的脑部MRI图像超分辨重建
Super-Resolution Reconstruction of Brain MRI Images Based on Differential Curvature Grouping Mixture Model
-
摘要: 核磁共振成像(magnetic resonance imaging, MRI)能够提供丰富的病理信息, 在脑损伤的诊断和治疗中具有重要意义, 受采样时间和现有医疗设备的限制, 临床上很难获得高分辨率的 MRI 图像. 为此, 提出一种基于差分曲率分组混合模型的超分辨重建方法. 首先在梯度特征提取的基础上引入差分曲率算法, 进一步检测图像的边缘、斜坡等特征结构, 并将特征块分为平滑区域、纹理区域和边缘区域 3 组; 然后基于学生 t 分布混合模型分别学习 3 组特征区域的模型参数; 最后选取多个似然概率较大的子分布共同重建高分辨率图像块. 在癌症成像档案库数据集上的实验结果表明, 在×2,×3 和×4 超分辨任务下, 所提方法的平均峰值信噪比分别为 41.36 dB, 35.01 dB 和 31.32 dB, 平均结构相似度分别为 0.984 8, 0.941 5 和 0.879 5; 与现有的超分辨重建方法相比, 该方法重建的 MRI 图像纹理细节更丰富、边缘更清晰, 并且重建时间更短.Abstract: Magnetic resonance imaging (MRI) provides rich pathological information which is of great significance in diagnosis and treatment of brain lesions. High resolution MRI images are hard to obtain in clinic due to the limitations of sampling time and existing medical equipment. To address these problems, a super-resolution(SR) reconstruction method is proposed based on differential curvature grouping mixture model. Firstly, a differential curvature algorithm is introduced on the basis of gradient feature extraction to detect edges, slopes and other feature structures of the image. Following that, the feature blocks are divided into three groups including smooth, texture and edge regions. Secondly, the student t-distribution mixed model is applied to learn the model parameters of the three sets of feature regions. Finally, multiple distribution models with larger likelihood probability are selected to reconstruct high-resolution image patches. The experiments on the cancer imaging archive (TCIA) dataset show that this method achieves an average peak signal-to-noise ratio (PSNR) of 41.36 dB, 35.01 dB and 31.32 dB with an average structural similarity index (SSIM) of 0.984 8, 0.941 5 and 0.879 5 respectively for ×2,×3 and×4 SR. Compared with some current SR reconstruction approaches, the proposed method reconstructs more reasonable images at the cost of less time with richer texture details and clearer edges.