부산대학교 도서관

Fast and quantitative image reconstruction for a new dedicated brain PET camera, CareMiBrain, is presented using the open-source Python package NiftyPET. The camera consists of 48 monolithic LYSO crystals arranged in 3 rings of 16 detector modules each, with an effective 240 mm transaxial FOV and 152 mm axial FOV and resolution below 2 mm. The coordinates of any photon detection at each module are encoded using 96 96 virtual pixels, which are then written to list-mode (LM) data×for coincidence and single events. The image generation pipeline, from LM data processing to image reconstruction is performed on graphics processing units (GPU) using NiftyPET: First, the prompt LM data is processed producing 5184 sinograms in span-1 with 192 projection angles and 240 bins, while single events are used to estimate random event sinograms; the centre of mass of the radio-distribution in the projection space is used for motion detection with a temporal resolution of 1 second. The normalisation is performed using a scatter-free long acquisition of a dedicated ring phantom. Scatter correction is performed using a fully 3D voxel-driven scatter model (VSM); forward and back projections are calculated on the fly using the ray-driven Siddon algorithm. The above computations are performed using high-throughput GPU routines while enabling easy access for data quality checks at any point of the image generation pipeline. The quantitative performance was evaluated using the Derenzo and uniform cylindrical phantoms, demonstrating accurate corrections for photon attenuation, scatter and random events across a range of radioactivity doses.