Fig. 6
From: Deep learning for Dixon MRI-based attenuation correction in PET/MRI of head and neck cancer patients
The worst performing cases, based on tumor (green delineation) SUVmean error, for (A) the deep learning method (error of −4.1% for PETDeep and −6.7% for PETAtlas) and for (B) the atlas-based method (error of −0.8% for PETDeep and −14.4% for PETAtlas). (A) Axial and sagittal slices of a 68 year old female with cancer of the right tonsil and bilateral lymph node involvement (T3N2M0). The large single void in the MRI affects both MR-AC methods. Notice, that the deep learning method partially adds tissue within the MRI signal void, and that the atlas-based method adds the jawbone despite the missing signal. (B) Axial and coronal slices 42 year old male with cancer of the left tonsil (T1N1M0). The vendor-provided atlas-based method is affected by a fat–water swap (fat becomes soft tissue and vice versa) and some of the air in trachea is segmented as soft tissue (axial slice)