Click here to

Session: Machine Learning for Adaptive Radiotherapy [Return to Session]

Integration of Attention-GAN to Predict Mid-Treatment Dual-Energy CT for Head and Neck Adaptive Radiation Therapy

Y Yan1,2*, R BAYLISS1, H Emami Gohari3, M Dong3, C Glide-Hurst1,2, (1) Department of Human Oncology, University of Wisconsin-Madison, Madison, WI, (2) Department of Medical Physics, University of Wisconsin-Madison, Madison, WI, (3) Department of Computer Science, Wayne State University, Detroit, MI


MO-H345-IePD-F2-1 (Monday, 7/11/2022) 3:45 PM - 4:15 PM [Eastern Time (GMT-4)]

Exhibit Hall | Forum 2

Purpose: Adaptive radiation therapy (ART) in head and neck cancer (HNC) enables higher therapeutic ratios and enhanced organ at risk (OAR) sparing at the expense of requiring significant resources to support replanning. We propose incorporating a novel attention-guided generative adversarial network (Attention-GAN) to predict mid-treatment high quality dual-energy computed tomography (DECT) datasets, with the overall goal of creating initial treatment plans robust to on-treatment changes.

Methods: Two DECTs (pre-treatment simulation and 21-62 days for mid-course timepoint) from 12 HNC patients were used to train the Attention-GAN. Key components include the generator (9-block residual network (ResNet) predicting mid-course DECTs from primary DECTs) and a discriminator (5-layer convolutional neural network (CNN) to discriminate real mid-course DECTs from predicted). Attention guidance was implemented in the intermediate layer of the discriminator to draw the attention of the generator to regions deviating substantially between the two DECTs. Model performance between real and predicted mid-treatment DECTs was evaluated using mean absolute error (MAE), structural similarity (SSIM), and radiomics features.

Results: MAEs between real and predicted mid-course DECTs were 176.3±66.9 HU over the entire field of view, 108.2±40.3 HU for parotids, 148.4±143.0 HU for gross tumor volume (GTV) and 45.4±12.3 HU for lymph nodes. Full FOV SSIMs between real and predicted mid-course DECTs were 0.95±0.03. Comparable radiomics features were obtained between the real and simulated mid-course DECT. Qualitative review of attention maps suggests that deviation in patient positioning, image co-registration and patient anatomy all contribute to spatial attention which guides Attention-GAN training process.

Conclusion: Attention-GAN was tested and evaluated for mid-course DECT prediction task to support ART. While the overall MAE and SSIM were promising, these preliminary results suggest non-anatomical factors also influence the training process. Future work will include a structure-guided strategy to better handle changing anatomy between timepoints and improve overall mid-treatment prediction performance.

Funding Support, Disclosures, and Conflict of Interest: Research collaborations with Philips Healthcare, GE Healthcare, ViewRay, Inc., and Modus Medical. Research partially supported by the National Cancer Institute of the National Institutes of Health under Award Number R01CA204189 and R01HL153720.


Radiation Therapy, Dual-energy Imaging, Computer Vision


IM/TH- Image Analysis (Single Modality or Multi-Modality): Computer/machine vision

Contact Email