Click here to

Session: Multi-Disciplinary BLUE RIBBON [Return to Session]

CBCT Forecasting - A Convolution-LSTM Based Machine Learning Framework for Predicting “ahead-Of-Time” Patient Anatomy, Dose to Be Delivered, and Dose Trend in Head and Neck Radiotherapy

A Santhanam1*, B Stiehl1, M Lauria1, R Savjani1, S Gros2,D Low1, (1) University of California, Los Angeles, Los Angeles, CA, (2) Loyola University Chicago, Chicago IL


SU-I400-BReP-F2-4 (Sunday, 7/10/2022) 4:00 PM - 5:00 PM [Eastern Time (GMT-4)]

Exhibit Hall | Forum 2

Purpose: In head and neck (H&N) radiotherapy, monitoring daily treatment dose can avoid overdosing organs-at-risk (OARs). Since re-planning steps require time, it is important to predict “ahead-of-time” the fraction at which an OAR overdose may occur. In this abstract, we present a novel forecasting framework that predicts the patient anatomy for the next fraction given a set of consecutive daily CBCT images.

Methods: We developed a convolution LSTM based machine learning framework for predicting the patient anatomy for the next fraction. A set of 20 head and neck patients with 30 fractions are retrospectively analyzed. Using RTapp software, the daily anatomical changes in the simulation CT were computed using daily CBCTs and the subsequent delivered dose was computed. The anatomical changes in the sim-CT were then exported for training purposes. The machine learning framework focused on predicting the next fraction’s geometry as a function of the series of images acquired before this fraction. We employed a convolution-LSTM framework, which combines the Long Short Term Memory (LSTM) with the convolution neural network architectures. A set of four sequential convolution-LSTM layers were considered. The results are then combined using a custom-built n-dimensional convolution layer in order to predict the 3D frame. For training purposes, we considered the patient anatomy in the first 10 fractions as input and predicted the H&N geometry for the next fraction.

Results: The training accuracy was observed to 82% demonstrating a good predictive capability. The validation accuracy was close to 75% demonstrating that the systematic changes in the H&N anatomy was effectively tracked and predicted. The dose prediction was observed to have a 94% accuracy showing that a good per-fraction treatment dose can be predicted ahead-of-time.

Conclusion: This work demonstrates that treatment adaptation can be achieved by predicting fractions when an overdose or underdose can occur.

Funding Support, Disclosures, and Conflict of Interest: This work is supported by the Tobacco Related Disease Research Program 27IR 0056, NIH R56 1R56HL139767 01A1, Ken and Wendy Ruby Foundation, and the UCLA Department of Radiation Oncology


Motion Artifacts, Patient Movement, Targeted Radiotherapy


Not Applicable / None Entered.

Contact Email