Semi-Automatic Synthetic Computed Tomography Generation for Abdomens Using Transfer Learning and Semi-Supervised Classification

摘要

We proposed a new semi-automatic method for generating synthetic CT images from modified Dixon (mDixon) MR data of abdomens to address the challenges of PET/MR attenuation correction. To avoid both qualitative and quantitative PET errors that could compromise diagnostic accuracy, attenuation correction is necessary. A relatively robust existing method, MR-based synthetic CT generation, also requires advantaged MR sequences, which are technically challenging. To address this problem, we proposed a semi-automatic method that only requires mDixon MR sequences and generates synthetic CT images using machine learning methods. i.e., transfer learning and semi-supervised classification (SA-SC-TS). The significance of our efforts can be summarized into three points. (1) Our method generates synthetic CT images from challenging abdomen images using only the mDixon sequence. (2) With the combination of transfer learning and semi-supervised classification, SASC-TS can partition the voxels of. MR images into four classes corresponding to four types of tissues (i.e., fat, soft tissue, air and bone) and generate synthetic CT images based on this result (3) Benefitting from the semi-supervised classification, only a small amount of supervised information is needed in SA-SC-TS, which reduces the time consumption for radiologists to mark MR images. The experimental results indicate that the proposed SA-SC-TS method can effectively generate synthetic CT images from challenging abdomen images using mDixon MR sequence data only.