MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Deep Learning to Predict Dosimetric Metabolic Response Map Using Longitudinal 18F-FDG PET/CT Images for Pancreatic Cancer Patients

Y Yue1*, K Huang1 , P Maxim1 , S Ellsworth1 , R Tuli2 , (1) Indiana University- School of Medicine, Indianapolis, IN(2) Memorial Sloan-Kettering Cancer Center, New York, NY

Presentations

(Sunday, 7/14/2019) 5:00 PM - 6:00 PM

Room: 221AB

Purpose: To predict to voxel-wise therapy response for pancreatic radiotherapy utilizing pre-radiotherapy PET/CT and planned dose, to achieve dose escalation and duodenum avoidance based on the predicted response/risk map toward response-guided adaptive radiotherapy.

Methods: 18FDG-PET/CT scans were obtained from 21 locally advanced pancreatic cancer patients, 2-week prior to (pre-RT), 12-week following treatment (post-RT). All PET/CT images were registered to the planning CT and planned dose. PTV and surrounding duodenum were selected as the volume-of-interest (VOI), dividing into 4.8x4.8x4.8mm^3 subvolumes. Local response was characterized in PET images as the standard uptake value (SUV) changes larger than a threshold, which was the mean uptakes difference between pre- and post-RT in un-irradiated tissue. The voxel-wise dosimetric-metabolic response were established based on mean dose and mean metabolic activities. A 50-layer deep residual network (ResNet-50) was implemented to predict the metabolic response using pre-RT PET and dose, where the model was pre-trained over a one million natural images database, ImageNet. A transfer learning technique was implemented to fine-tune the model using the patient data. 16 patients were used to train the model, and data of 5 patients were used as test data set. The performance of prediction was evaluated by classification F-score.

Results: Follow-up scans of 21 patients show significant response to treatment in terms of changes of SUVmax (p=0.031). In ResNet transfer learning, over 9600 subvolumes from 16 patients were randomly selected as training data, and 2200 subvolumes were used as the validation set. The training loss is 0.253(Accuracy:0.906), validation loss 0.377(Accuracy:0.804). The results of predictions for 5 patients show F-score is 83.9% with recall=81.5%, precision=86.4%.

Conclusion: Deep learning prediction of local metabolic response provides a feasible approach for evaluating efficacy of local control of pancreatic cancer treatment. The results can be used to identify local failure of disease, toward implementing response-driven adaptive radiotherapy.

Keywords

Image Analysis, PET, Dose Response

Taxonomy

TH- response assessment : PET imaging-based

Contact Email