A Two-Stage Whole Body Bone SPECT Scan Image Inpainting Algorithm for Residual Urine Artifacts Based on Contextual Attention

Whole body bone SPECT scan is an important inspection method to detect early malignant tumor or evaluate bone metastases. As the specificity of the SPECT images is significantly affected by noise, metal artifacts, or residual urine, which may mislead the doctor’s diagnosis. To address this issue, a two-stage whole body bone SPECT scan image inpainting algorithm for residual urine artifact based on contextual attention is proposed in this paper. In the first stage, TransUNet framework is utilized to segment the whole bone SPECT images according to the location of the residual urine. In the second stage, the artifact in the segmented image is recovered by a contextual attention network, which can effectively deal with residual urine artifact and improve the quality of the recovered image. Besides, an extended image dataset named as EX-BS-90K consisting more than 90k whole body bone SPECT scan images is presented in this paper and used to train the proposed two-stage inpainting model. The PSNR and SSIM are calculated as evaluation metrics of the proposed algorithm and the experimental results demonstrates that the proposed method can effectively inpaint the original bone structure in the artifact region, which can increase the specificity of the SPECT images. Meanwhile, the proposed method are compared with recent works and the comparison results demonstrates that our method provides better solution for whole body bone SPECT images inpainting problems. The code and dataset are available: https://github.com/Zhoupixiang/Two-stage-BS_inpainting .