Click here to

Session: Outcome Modeling and Assessment [Return to Session]

Compact Neural Network to Predict Radiation-Induced Lymphocyte Depletion

Y Kim1*, I Chamseddine2, A Saraf3, Y Cho4, F Keane5, W Sung6, H Yoon7, H Paganetti8, J Kim9, S Cho10, C Grassberger11, (1) KAIST, Daejeon, ,KR, (2) Harvard Medical School/Massachusetts General Hospital, Quincy, MA, (3) Massachusetts General Hospital, ,,(4) Yonsei University College of Medicine, Seoul, ,KR, (5) Massachusetts General Hospital, ,,(6) The Catholic University of Korea, Seoul, ,KR, (7) Yonsei University College of Medicine, Seoul, ,KR, (8) Mass General Hospital and Harvard Medical School, Boston, MA, (9) Yonsei University College of Medicine, Seoul, ,KR, (10) KAIST, Daejon, ,KR, (11) Massachusetts General Hospital, Cambridge, MA


TU-A-TRACK 6-1 (Tuesday, 7/27/2021) 10:30 AM - 11:30 AM [Eastern Time (GMT-4)]

Purpose: Radiation-induced lymphopenia (RIL) is known to be associated with survival after radiotherapy (RT). We investigate the correlation of dose distributions with RIL in lung and liver cancer patients and develop a compact neural network to predict RIL in this study.

Methods: We investigated dose distributions in 84 lung cancer patients and 63 liver cancer patients treated at two institutions, and investigated RIL directly after RT and the lymphocyte recovery 3 months later. Dose distributions were deformably registered to a common target coordinate system for spatial normalization, and non-parametric correlation tests were used to study local effects of dose to specific regions. A compact multi-path neural network was designed that integrates clinical patient data in fully connected layers with the dose distribution in convolutional layers for predicting the post-RT absolute lymphocyte count of lung cancer patients. We used relatively a small (less than 3) number of convolutional layers, a limited number of kernels, and extensive max-pooling for a light-weight network. Training and validation were performed separately on data from different institutions.

Results: Clear correlation of RIL with dose was observed around large vessels - pulmonary vessels and heart in lung cancer patients and large hepatic vessels in liver cancer patients. When focusing on lymphocyte recovery, these correlations shifted to bone marrow. Keeping the neural network structure compact enabled convergence of the network in both training (n=63) and validation sets (n=21), and the network predicted the post-RT lymphocyte count (average 0.9/μl, range [0.1-2.2] /μl) with a mean square error of 0.06-0.08.

Conclusion: Dose to large vessels is strongly correlated with RIL directly after RT in liver and lung cancer patients, while dose to bone marrow correlates with the recovery from RIL. The developed neural network integrates 3D dose information directly with clinical data and successfully predicts RIL.

Funding Support, Disclosures, and Conflict of Interest: This study is supported by NCI (grant nr: R21 CA241918) and Korean Health Industry Development Institute (grant number:HI19C1332)



    NTCP, Dose Response, Blood Vessels


    Not Applicable / None Entered.

    Contact Email