Development of a new, neural network-based partial volume correction method for SPECT/CT for an improved quantification of organ doses in molecular radiotherapies

• Applicant: Dr. Johannes Tran-Gia
• Subject Area: Medical Physics, Biomedical Technology ; Nuclear Medicine, Radiotherapy, Radiobiology
• Term: since 2020
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 451001990

In molecular radiotherapies, a radiopharmaceutical, mostly labelled with a beta-emitter, is administered to a patient. While radiopharmaceuticals, which are frequently injected systemically, are effective at destroying malignant tissue, they also affect healthy tissue. Therefore, patient-specific dosimetry is essential for predicting the safety of a treatment and for verifying the therapy response. Establishing a reliable dose-response relationship involves several steps. The first is the determination of the activity of the radiopharmaceutical taken up by the organ of interest over time. As uncertainties propagate along each of the subsequent steps (integration of the time-activity curve, absorbed dose calculation), a reliable activity quantification is essential.For measuring peri-therapeutic activity distributions, quantitative single photon emission computed tomography in combination with computed tomography (SPECT/CT) is typically applied. After a tomographic reconstruction including the most common corrections (e.g., for attenuation, scatter, and collimator-detector response), the counts detected by the imaging system are converted to activity concentration. After applying all available corrections and a traceable count-to-activity conversion, however, a resolution-related blur remains. This so-called partial volume effect represents a major challenge that must be solved to enable a reliable absorbed dose calculation. Another source of uncertainty is the frequently insufficient validation of quantitative imaging setups. Calibration is mostly performed based on commercially available, non-patient-specific phantoms of simple geometries (e.g., spheres or cylinders). As partial volume errors are dependent on the spatial distribution of the activity in the entire field-of-view, a validation on more patient-specific phantoms would substantially improve confidence in quantitative SPECT/CT.A major goal of the proposed project is the development of a new method for partial volume correction (PVC) based on artificial intelligence, more specifically on convolutional neural networks (CNNs). To generate the large amount of data required for training CNNs, Monte Carlo simulations of SPECT/CT will be applied in collaboration with the University of Lund’s Nuclear Medicine Physics group, one of the world leading experts in this field. To overcome the validation difficulties mentioned above, a patient-specific kidney phantom will be designed based on morphological imaging data of a patient. Fillable phantoms of this design will be fabricated using two of the most common 3D printing technologies, fused-deposition modeling and stereolithography for identifying and optimizing the most promising method. SPECT/CT measurements of the patient-specific phantoms will be performed for validation of the newly developed PVC correction method. Finally, the newly developed algorithm will be tested in a small patient cohort for assessing the clinical applicability.

• DFG Programme: Research Grants