Click here to

Session: Therapy General ePoster Viewing [Return to Session]

Improved Dose Prediction Performance with Data Clustering and Transfer Learning On H&N Cancer Patients

D Nguyen*, T Bai, A Sadeghnejad Barkousaraie, R McBeth, A Balagopal, S Jiang, Medical Artificial Intelligence and Automation (MAIA) Laboratory, UT Southwestern Medical Center, Dallas, TX, USA


PO-GePV-T-283 (Sunday, 7/25/2021)   [Eastern Time (GMT-4)]

Purpose: We propose a deep learning (DL)-based clustering and Transfer Learning technique for improved model performance and test it on dose prediction of head and neck (H&N) cancer patients.

Methods: We used H&N cancer patient data available from the OpenKBP Challenge, consisting of 200 training, 40 validation, and 100 testing. This includes CTs, planning target volumes (PTV) and organs-at-risk (OAR) masks, 3D dose distributions. The dose distribution was generated from a 9 beam equidistant IMRT setup with 6 MV fields. As baseline, we trained a U-net style architecture model on all of the training data to take the patient anatomy as input to predict the dose distribution. For the clustering method, we first utilized an encoder-decoder style model and trained it to take the anatomy as input to predict the dose distribution, and then used the encoder to transform all the data into the latent space. A TSNE calculation is made on the data in the latent space, and clusters of data are determined. Models of the same U-net style architecture as the baseline are then trained to predict the dose distribution in one of two ways: 1) training from scratch on each data cluster or 2) Transfer Learning (TL) from the baseline model to each data cluster.

Results: Compared to the ground truth data, the Cluster+TL method performed the best in regard to minimizing the error in structure mean and max doses, dose coverage, and isodose similarities between its prediction and the ground truth dose data. Training from scratch on each cluster without TL performed similarly to the baseline model likely due to tradeoff between data quantity and precision.

Conclusion: We show that our method of Clustering+TL can further improve the result from baseline. Clustering without TL may require more data before it can outperform the baseline model.

Funding Support, Disclosures, and Conflict of Interest: This study was supported by the National Institutes of Health (NIH) R01CA237269 and the Cancer Prevention & Research Institute of Texas (CPRIT) IIRA RP150485.



    Radiation Therapy, Dose


    TH- External Beam- Photons: Treatment planning using machine learning/Knowledge Based Planning/automation

    Contact Email