Please use this identifier to cite or link to this item:
Title: The Nephroblastoma Oncosimulator: Clinical adaptation for patients of distinct histologic profiles using high performance computing
Authors: Παναγιωτίδου, Φωτεινή
Σούντρης Δημήτριος
Keywords: Oncosimulator
Nephroblastoma Oncosimulator
Wilms’ tumor
In silico medicine
Multiscale cancer modeling
Digital twin
High performance computing
Virtual patients
Clinical adaptation
Issue Date: 19-Oct-2021
Abstract: The Oncosimulator is a top-down discrete entity-discrete event model for multiscale cancer modeling, that represents the tumor in silico as a 3D matrix of voxels and utilizes the "summarize and jump" strategy for the simulation of its evolution on the cellular and tissue level of biocomplexity. The Nephroblastoma Oncosimulator is a version of the Oncosimulator software that is specific to the Wilms’ tumor, which is the most common kidney tumor in childhood. It was developed by members of the In Silico Oncology and In Silico Medicine Group, ICCS, ECE, National Technical University of Athens, under the lead of G. Stamatakos and the clinical advisorship of N. Graf, University of Saarland, from 2006 to 2019. The ultimate goal of the Oncosimulator and its neoplastic disease specific versions is to act as medical advisors for personalized medicine by creating a digital twin for the tumor and its normal tissue microenvironment. However, for that purpose the patient-specific medical data need to be translated in silico and, even though this is a trivial task when the imaging data and the size and shape of the tumor are considered, the values for the simulation input parameters that define the simulated biologic processes are not unambiguously derived. The purpose of the present work is to address this issue by exploring high performance computing methods for the performance enhancement of a single simulation and the efficient performance of a multitude of virtual patient simulations for each physical patient, with each virtual patient having a distinct value for the joint distribution of the simulation input parameters. The cluster of virtual patient simulations were executed for three already treated patients of distinct histologic profiles and corresponding risk groups, whose medical data were provided by the Saarland University Hospital, with each virtual patient being assigned a distinct value for the joint distribution of the simulation input parameters and the cell kill ratio parameter, which expresses the effects of the administered therapy, being explored for each virtual patient until the simulation outcomes matched the actual patient outcomes. Prior to the execution of the cluster of virtual patient executions that implement the in silico clinical adaptation of the CKR parameter, the size of the voxel in the 3D tumor matrix was adjusted for the inputs that correspond to each dataset patient in the context of a data preprocessing step that addresses the simulation resolution and costs tradeoff problem, code optimizations were applied for the utilization of improved resources and the concurrent execution of computations when the data dependencies allow it and sensitivity analysis was performed for the verification of the preprocessed input data and the optimized source code via the comparison to predecessor verified simulation executions. The data preprocessing step introduces significant speedup in the simulation execution time (approximately 37% - 688%) and memory footprint reduction of the simulation execution (approximately 87% - 94%), while indirectly improving the speedup potential of the code optimization step according to Amdhal’s law by rendering the simulation workload balance more fair and less intense on the parts that do not benefit from the improved resources. The source code optimization introduces a speedup of approximately 16% - 24% by focusing on performing concurrent computations for the data dependencies free third simulation scan that concludes the simulation on the cellular level of biocomplexity. The final clinical adaptation step was perfomed for 20 and 200 virtual patients for each dataset patient, producing a distinct distribution, i.e. mean value and standard deviation, for the CKR parameter for each distinct risk group of the patient dataset.
Appears in Collections:Διπλωματικές Εργασίες - Theses

Files in This Item:
File Description SizeFormat 
Thesis.pdf31.11 MBAdobe PDFView/Open

Items in Artemis are protected by copyright, with all rights reserved, unless otherwise indicated.