Cross-modality Synthesis from MRI to PET Using Adversarial U-Net with Different Normalization

Multi-modality biomedical images (especially magnetic resonance imaging (MRI) and positron emission tomography (PET)) are critical for auxiliary diagnosis in brain diseases. However, there are some common concerns in PET scans including the high cost, the usage of radioactive tracer and so on. These factors jointly result in a lack of PET datas. To overcome this limitation, in this study we proposed an effective U-Net architecture with the adversarial training strategy to synthesizing PET datas from their corresponding MRI. In addition, we notice that many prior works commonly adopted Batch Normalization (BN) as the normalization method by default. But in this specific task which always needs a small mini-batch, BN's performance decreases rapidly. To alleviate this issue, we compared the performance of using other popular normalization methods rather than using BN as default. Experimental results on a subset of ADNI database demonstrated that the synthetic PET images from our proposed method were reasonable, and it was valuable to replace the Batch Normalization with the Instance Normalization in the tasks of cross-modality synthesis. Our study could help future research in this field.