Feasibility of Dynamic SPECT-Renography with Automated Evaluation Using a Deep Neural Network

Detta är en Master-uppsats från Lunds universitet/Sjukhusfysikerutbildningen

Författare: Viktor Rogowski; [2021]

Nyckelord: Medicine and Health Sciences;

Sammanfattning: Introduction: Renography is a standard diagnostic examination that evaluates renal function, renal pelvic dilatation and urinary obstruction. Renography is performed by injecting a radiopharmaceutical (predominately 99mTc-MAG3) and using gamma camera to image the biodistribution in a dynamic sequence. Evaluating renography images encompasses many difficulties regarding background correction for activity in blood and tissue located anterior and posterior of the kidney. With obstructed renal function activity accumulates in the liver and border of the kidneys is difficult to distinguish. A possible solution to these difficulties is the novel multi-detector Single Photon Emission Computed Tomography (SPECT) systems which, unlike conventional systems with which only a few projections are measured simultaneously, allows for a simultaneous 360-degree measurement of the biodistribution and thereby enables dynamic SPECT. However, with the new system, other problems arise regarding image noise and larger workload on the technologist with evaluation. By using convolutional neural network (CNN) to do automatic segmentation these obstacles can be avoided. However, to train a CNN thousands of images are required which can be obtained with Monte Carlo simulations. The aim of this thesis is to evaluate the feasibility of four dimensional (4D)-renography and compare it to conventional renography and to train a CNN for three dimensional (3D) semantic segmentation of 4D-renography images with the use of Monte Carlo simulated images and to evaluate the model. Material and Methods: 15 4D anthropomorphic digital phantoms (XCAT) with 15 different sets of renal functions and split renal functions were simulated with the Monte Carlo program SIMIND to mimic the VERITON Cadmium Zinc Telluride (CZT) camera. Simulations were set to 360-degree rotation and 120 projections. Each organ was simulated individually, and ground truth masks of the organs were created from the phantoms. The simulations were assembled into 3 sets of data: two dimensional (2D) planar images with posterior projection, which resemble the images corresponding to conventional renography studies, geometric mean corrected images and 3D dynamic tomographic projections. The assembled images were normalized to a realistic activity level, acquisition time, sensitivity correction and Poisson distributed noise was then added. Tomographic projections were further reconstructed using an in-house OS-EM method. Evaluation was performed with the use of simulated masks and with extracted count rate a split renal function is calculated and compared between the different sets. A CNN was trained by using 3D U-net network. Tomographic images were modified resulting in 24894 unique training images and masks. Evaluation of split renal function using CNN segmentation was compared to planar evaluation. Results: The tomographic evaluation showed less deviation from the true value in the majority of cases when compared to posterior planar evaluation. For lower renal function with greater difference in split renal function planar evaluation showed greater deviation from true value. Geometric mean image evaluation showed greater deviation for low renal function and generally worse than both tomographic and posterior planar evaluation. Evaluation using neural network showed less deviation for most cases compared to posterior planar evaluation. Conclusion: The results show that 4D-renography may be feasible, with a possible accuracy of a split renal function measurement that is better than the conventional methods. Scaling the collimator sensitivity to the VERITON cameras sensitivity results in noisy images which limits the minimum possible voxel size and time resolution. Evaluation with the CNN shows great promise. Further development and evaluation of clinical images is needed to ensure accuracy.

  HÄR KAN DU HÄMTA UPPSATSEN I FULLTEXT. (följ länken till nästa sida)