Loading...

Personalized computational human phantoms via a hybrid model-based deep learning method

Khodajou Chokami, H ; Sharif University of Technology | 2020

940 Viewed
  1. Type of Document: Article
  2. DOI: 10.1109/MeMeA49120.2020.9137114
  3. Publisher: Institute of Electrical and Electronics Engineers Inc , 2020
  4. Abstract:
  5. Computed tomography (CT) simulators are versatile tools for scanning protocol evaluation, optimization of geometrical design parameters, assessment of image reconstruction algorithms, and evaluation of the impact of future innovations attempting to improve the performance of CT scanners. Computational human phantoms (CHPs) play a key role in simulators for the radiation dosimetry and assessment of image quality tasks in the medical x-ray systems. Since the construction of patient-specific CHPs can be both difficult and time-consuming, nominal standard/reference CHPs have been established, yielding significant discrepancies in the special design and optimization demands of patient dose and imaging protocols for most medical applications. Therefore, the aim of this work was to develop a personalized Monte-Carlo (MC) CT simulator equipped with a fast and well-structured tool-kit called DeepSegNet for automatic generation of patient-specific CHPs based on MRI images, working under two principal algorithms. To this end, we first developed a 3D convolutional neural network (3DCNN) for the automated segmentation of 3D MRI images to detect anatomical organs/tissues. Then, a 3D voxel merging (3DVM) algorithm constructing CHPs and making fast MC calculations were developed. The proposed 3DCNN benefits from the main merit of residual networks by designing a 15-layer model. Next, the 3DVM algorithm utilizes the segmented data acquired from the former step, to create realistic and optimized CHPs by material mapping and voxel size manipulating. The performance of our 3DCNN model on 20 patients as test cases was 84.54% and 74.52% in terms of average accuracy and Dice-Coefficient, respectively, outperforming SegNet, as a comparable method by 2%. Finally, we developed an MC CT simulator by implementing a set of our generated CHPs. The efficiency of our 3DVM algorithm in constructing CHPs was assessed in terms of MC execution time and the number of merged voxels representing occupied storage memory and compared to the existing lattice method. Besides, the accuracy of our 3DVM investigated through the estimation of patient dose maps and image reconstruction. Results demonstrated a significant reduction of about 96% in the number of voxels and a 15% reduction in MC execution time for x-ray photon transportation while keeping the same accuracy. Therefore, this software package has a strong potential in the optimization of therapeutic and radiological imaging procedures. © 2020 IEEE
  6. Keywords:
  7. Computational human phantoms ; Follow-up CT imaging ; Lattice method ; Monte-carlo method ; Semantic segmentation ; Voxel merging algorithm ; Computerized tomography ; Convolutional neural networks ; Digital storage ; Image enhancement ; Image segmentation ; Learning systems ; Magnetic resonance imaging ; Medical applications ; Medical imaging ; Phantoms ; Simulators ; X rays ; Automated segmentation ; Automatic Generation ; Design and optimization ; Future innovations ; Geometrical designs ; Image reconstruction algorithm ; Protocol evaluation ; Radiological imaging ; Deep learning
  8. Source: 15th IEEE International Symposium on Medical Measurements and Applications, MeMeA 2020, 1 June 2020 through 3 June 2020 ; July , 2020
  9. URL: https://ieeexplore.ieee.org/document/9137114