Training of neural network for attenuation correction in PET/CT
A method is provided for generating an attenuation map for PET image reconstruction. The method includes training a deep convolutional neural network (DCNN) model by minimizing a loss function between initial input image data and the attenuation map generated by a spectral CT scan as supervised data. Further, the method includes obtaining PET data from a scan of a subject and reconstructing a PET image from the PET data and an attenuation map output from the DCNN. The initial input image data can be from a conventional CT scan with or without beam-hardening correction or the input image data can be from a phantom, a simulation, or a SPECT image.
1 . A method, comprising:
receiving integration computed tomography (CT) image data of a subject, the integration CT image data being obtained by a non-spectral CT scan; and
generating an attenuation map for Positron Emission Tomography (PET) image reconstruction by inputting the received integration CT image data into a trained deep convolutional neural network (DCNN) model that outputs the attenuation map, wherein the DCNN model was trained using integration CT image data obtained by a non-spectral CT scan and corresponding training attenuation map data generated from spectral CT image data obtained by a spectral CT scan.
2 . The method of claim 1 , further comprising:
obtaining PET data from a scan of a subject; and
reconstructing a PET image from the obtained PET data and the generated attenuation map.
3 . The method of claim 1 , further comprising obtaining the training attenuation map data from input image data obtained from a first CT scan having a first energy below 90 kVp and from a second CT scan having an energy above 125 kVp.
4 . The method of claim 3 , further comprising:
generating a first basis material image and a second basis material image based on the input image data, and
generating the attenuation map from the first and second basis material images.
5 . The method of claim 1 , further comprising obtaining the training attenuation map data from input image data, which is produced by a spectral CT scan.
6 . The method of claim 1 , further comprising obtaining the training attenuation map data from input image data, which is obtained from a fast kV switching dual-energy CT scan.
7 . The method of claim 1 , further comprising obtaining the training attenuation map data from input image data, which is obtained from a scan using a photon-counting CT apparatus.
8 . The method of claim 1 , wherein the received CT image data was generated with beam-hardening correction turned off.
9 . The method of claim 1 , further comprising correcting the received CT image data using a beam-hardening correction process prior to inputting the CT image data into the trained DCNN model.
10 . The method of claim 1 , wherein the received CT image data includes both single-energy CT image data having been beam-hardening corrected and single-energy CT image data not having been beam-hardening corrected.
11 . The method of claim 1 , further comprising training the DCNN model by minimizing a loss function.
12 . The method of claim 1 , further comprising obtaining the training attenuation map data from input image data, which is obtained from a simulation or from a digital phantom.