Abstract:
Our project aims to synthesize Positron Emission Tomography (PET) - Like images from a MRI scan from
artificial intelligence (AI) driven models. Dataset used in this regard is of 37 patients each having a T1w, FLAIR
and PET image modality. These images which were in Neuroimaging Informatics Technology Initiative (NIFTI)
format were pre-processed by converting into 2D tensors and extending them to 3D tensors by adding an extra
dimension. The T1 and FLAIR images are concatenated and given as input to pix2pix model while PET images
are set as the ground truth for our model. The synthesized output from the above model serves as the input to
another machine learning model which is a modified super resolution convolutional neural network (SRCNN)
called Fast Medical Image Super Resolution Method. This model maps a low resolution image to a super
resolution image thus giving us better images. For hardware implementation, FPGA and DSP Kit are utilized for
pattern recognition on the output PET-like image. Furthermore, the synthesis software model is uploaded on a
Raspberry Pi to allow for localization and environment integration. This project will aid in bridging the healthcare
gap by providing a non-invasive alternate for PET imaging by using easily accessible MRI data. It will also reduce
the need of costly PET scanners which are limited in Pakistan.
