MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Prediction of Post-Radiotherapy Residual Disease From Pre-Treatment Imaging Using Deep Learning

D Huff1*, T Bradshaw1 , R Jeraj1,2 , (1) University of Wisconsin-Madison, Madison, WI (2) University of Ljubljana, Ljubljana, Slovenia

Presentations

(Sunday, 7/29/2018) 3:00 PM - 3:30 PM

Room: Exhibit Hall | Forum 6

Purpose: Following radiotherapy (RT), small pockets of tumor can survive, leading to recurrent disease and poor outcomes. Identification of tumor subvolumes most likely to recur before treatment may enable treatment plan optimization and improved patient outcomes. Here, we assess the feasibility of a convolutional neural network (CNN)-based approach to the task of predicting residual disease location.

Methods: Fifteen canine subjects with sinonasal tumors received baseline CT, ¹�F-FDG, ¹�F-FLT, and �¹Cu-ATSM PET scans. Three to twelve months after radiotherapy (10 fractions, 42-50 Gy to GTV via Tomotherapy), residual disease was identified on ¹�F-FDG PET/CT, contoured using a level-set algorithm, and rigidly registered to pre-treatment images. A U-Net CNN model was implemented to predict post-RT residual disease from pre-RT imaging data, dose plans, and GTV masks. Model performance was evaluated with receiver operator characteristic (ROC) analysis and Dice coefficient between the true and predicted residual disease contours using leave-one-out cross-validation.

Results: The best model performance was achieved using input channels of CT, GTV, dose plan, and 18F-FLT PET images, with an ROC area under the curve (AUC) of 0.65±0.13 (mean±SD). Mean Dice coefficient at the optimal threshold for this method was 0.56±0.21 and ranged from 0.20 to 0.82 across patients. Models trained with FDG and Cu-ATSM PET images in place of FLT images achieved mean AUCs of 0.60±0.10 and 0.58±0.17 respectively, and mean Dice coefficients of 0.55±0.16 and 0.58±0.16, respectively. Patient-wise performance was moderately correlated across model input sets (Cu-ATSM vs. FLT: Spearman Ï?=0.65, Cu-ATSM vs. FDG: Ï?=0.64, FLT vs. FDG: Ï?=0.45), indicating that model performance is patient dependent.

Conclusion: The performance of a CNN-based approach for predicting post-treatment residual disease contours from pre-treatment images varied widely across patients. An expansion of patient population and further model optimization may improve performance and lead to a viable approach for identifying residual disease location.

Keywords

PET, Treatment Planning, Image-guided Therapy

Taxonomy

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

Contact Email