Click here to

Session: Multi-Disciplinary General ePoster Viewing [Return to Session]

Feasibility Study of Deep Learning Based Radiosensitivity Binary Classification Model Using Gene Expression Profiling

E Kim*, Y Chung, Hanyang University, Seoul, KR


PO-GePV-M-24 (Sunday, 7/25/2021)   [Eastern Time (GMT-4)]

Purpose: In radiotherapy, the prediction of radiosensitivity of tumors has been considered to be clinically relevant. In this prospect, the radiosensitivity prediction model was presented using deep learning in previous study, whose radiosensitivity were represented by survival fraction at 2 Gy (SF2). However, what clinically important is whether the cell is radiosensitive or radioresistant, rather than the SF2 value. This binary classification can support decision making of physician more efficiently and require fewer computational resources. With this perspective, we investigated the feasibility of the deep learning binary classification model of radiosensitive or radioresistant using gene expression profiling.

Methods: The gene expression profiling of the National Cancer Institute-60 tumor cell line was acquired from the gene expression omnibus database. SF2 values as a radiosensitivity indicator were obtained from previous publications, which were converted to one-hot encoded radiosensitivity as 0 if SF2>0.2 (radiosensitive) and 1 if SF2>0.8 (radioresistant). In this dataset, 36 triplicated data of 12 cell lines are included. The radiosensitivity prediction model was based on convolutional neural network, which consisted of 5 convolutional layers. The six fold cross-validation approach were applied to train and validate the model. The accuracy of the model was defined as a percentage of correctly classified samples.

Results: The classification accuracy was 97.22%. The only sample failed to classify correctly was third sample of MDAMB435 cell line, which were also failed to predict correctly in previous study. In terms of computing time, the binary model converged to the saturation point 10 times faster than the previous model for predicting the precise SF2.

Conclusion: This study suggests that the utilization of deep learning is feasible to develop the classification model deciding whether the cell is radiosensitive or radioresistant. Since the amount of data in this study was limited, further studies with additional data would be required.

Funding Support, Disclosures, and Conflict of Interest: This research was supported by Basic Science Research Program through the National Research Foundation of Korea (KRF) funded by the Ministry of Education (NRF-2018R1D1A1B07049228).



    Not Applicable / None Entered.


    Not Applicable / None Entered.

    Contact Email