Click here to

Session: Therapy General ePoster Viewing [Return to Session]

A Biology-Driven Deep-Learning Based Cell Survival Model for Proton Radiation

Y Lai1,2*, C Shen1, Y Chi2, F Guan3, X Jia1, (1) innovative Technology Of Radiotherapy Computation and Hardware (iTORCH) laboratory, Department of Radiation Oncology, University of Texas Southwestern Medical Center, Dallas, TX 75287, USA (2) Department of Physics, The University of Texas at Arlington, Arlington, TX 76019, USA (3) Departments of Radiation Physics, The University of Texas MD Anderson Cancer Center, Houston, Texas 77030, USA

Presentations

PO-GePV-T-444 (Sunday, 7/25/2021)   [Eastern Time (GMT-4)]

Purpose: It is challenging to construct a generic biophysical model with good generality to accurately predict cell survivals for proton radiation. This study proposes a Biology-guided Cell Survival deep neural network (BgCSDNN) to directly derive cell surviving fraction (SF) based on different types of DNA damages computed by microscopic Monte Carlo (MC) simulations.

Methods: GPU-based MC computation platform gMicroMC was used to simulate the physical and chemical stages of proton radiations and calculate different types of DNA damages following setup of a proton irradiation experiment on lung cancer H460 cells. Spectrum of DNA damage was calculated to account for the dose inhomogeneities among cells. A biology-driven neural network that predicts SF as a function of the DNA damage spectrum was constructed with the network structure reflecting major pathways of DNA damage repair. We performed an end-to-end training of the network using MC simulated DNA damage data with ten different LETs and tested the validity of the SF prediction model for two different LETs. Weights of pathways related to different DNA damage types in the trained network were analyzed.

Results: At 6 Gy, when LET increases from 0.9 to 19.0 keV/µm, double-strand breaks (DSBs) increases from 334±26.7 to 574±101.7 (mean ± standard deviation). The trained network achieved a mean relative error of 0.1 and 0.11 against experimental ground truth data on training data and test data, indicating generalizability of the model, which performed better than 0.23 for a LET-based linear-quadratic model. In the trained network, weights for DSB related pathways were ~0.4, substantially larger than the one for single strand break (~0.01), indicating the importance of DSBs in cell kills.

Conclusion: The novel BgCSDNN can accurately predict cell survival for proton radiation and provides biological insights to the cell survival process.

ePosters

    Keywords

    Not Applicable / None Entered.

    Taxonomy

    Not Applicable / None Entered.

    Contact Email

    Share: