MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel
  1. Home
  2. Meeting Program
  3. All Sessions
  4. All Multi-Disciplinary Scientific Session
  5. An Encoder-Decoder Based Convolutional Neural Network (ED-CNN) for PET Image Response Prediction Using Pre-RT Information: A Feasibility of Oropharynx Cancer IMRT

An Encoder-Decoder Based Convolutional Neural Network (ED-CNN) for PET Image Response Prediction Using Pre-RT Information: A Feasibility of Oropharynx Cancer IMRT

Y Chang1*, K Lafata2 , C Liu3 , C Wang4 , Y Cui5 , L Ren6 , X Li7 , Y Mowery8 , D Brizel9 , F Yin10 , (1) Duke University Medical Center, Durham, NC, (2) Duke University Medical Center, Durham, NC, (3) Duke Kunshan University, Suzhou, Jiangsu, (4) Duke University Medical Center, Durham, NC, (5) Duke University Medical Center, Durham, NC, (6) Duke University Medical Center, Cary, NC, (7) Duke University Medical Center, Durham, NC, (8) Duke University Medical Center, Durham, ,(9) Duke University Medical Center, Durham, ,(10) Duke University Medical Center, Durham, NC

Presentations

(Monday, 7/15/2019) 4:30 PM - 6:00 PM

Room: Stars at Night Ballroom 2-3

Purpose: To develop an encoder-decoder convolutional neural network (ED-CNN) for predicting treatment plan-specific PET image response using pre-RT information in oropharynx cancer cases undergoing IMRT.

Methods: This study incorporated 66 oropharynx cancer cases that received a pre-RT PET-CT scan and an intra-RT PET-CT scan after 24Gy IMRT delivery. Two scans were registered and resampled at same grid size. The model adopted 2D pre-RT PET, pre-RT CT, and planned dose distribution slices containing regions of interest (ROI) (GTV and oral cavity (organs at risk (OAR)) as input, and predicted PET image response (i.e., PET image changes from pre-RT PET). During model training, 60 cases were used to build an ED-CNN architecture: a 7-layer encoder and a 6-layer decoder were connected by 2 concatenated layers. Prediction in ROI was emphasized by a novel customized loss function, which was the sum of ROI-weighted mean square error and cross correlation. 10-fold cross-validation was implemented. In model testing, the predicted intra-RT PET of the other 6 cases were constructed based on the predicted PET response from pre-RT PET and compared with the ground-truth intra-RT PET. Cross correlation and gamma index passing rate (5% SUV difference/5mm distance-to-agreement) were used for quantitative evaluation.

Results: The predicted intra-RT PET agreed with the ground truth intra-RT PET on general contrast and shape of high-SUV region within ROI. The 3D cross correlation (meanSD) was 0.921±0.024/0.922±0.017 for GTV/OAR volume, and the 2D cross correlation was 0.940±0.022 for all testing axial slices. The 3D gamma passing rate was 0.963±0.019/0.921±0.024 for GTV/OAR volume, and the 2D gamma passing rate was 0.942±0.055 for all testing

Conclusion: Preliminary results suggest the feasibility of predicting intra-RT PET image response using pre-RT data. The proposed method demonstrated great potential for aiding treatment planning decision making such as selective dose de-escalation in oropharynx cancer IMRT.

Keywords

FDG PET, Dose Response, Image Analysis

Taxonomy

TH- response assessment : PET imaging-based

Contact Email

Additional Links