(Sunday, 7/29/2018) 2:05 PM - 3:00 PM
Room: Davidson Ballroom B
Purpose: This study aims to develop a novel method for CT ventilation imaging based on radiomic filtering and deep convolutional neural networks (CNNs).
Methods: Fifty-five quantitative radiomic features were considered as potential imaging biomarkers for pulmonary function. These features – which were chosen to collectively capture intensity variations, fine-texture, and course-texture within the lungs – were generalized to be a set of finite impulse response kernels. The first convolutional layer of a novel deep CNN was defined by spatially filtering each radiomic kernel throughout the lungs of CT images, producing a set of shift-invariant feature-maps. The feature-maps were then forward-propagated through 6 fully-connected hidden layers, a convolutional smoothing layer, and a regression layer connected to â?¶â?¸Ga-PET pulmonary ventilation measurements. Weight optimization was achieved via backpropagation with scaled conjugate gradient minimization. Following hyper-parameterization and sensitivity analysis, model performance was evaluated on 5 patients based on the correlation coefficient (Ï?), coefficient of determination (r²), and root mean square error (RMSE). Each model was first cross-validated by training on 2/3 of an image, and testing on 1/3. Model generalization was then more strictly evaluated by training and testing on different patient images. The Dice Similarity Coefficient (DSC) was used to quantify how well a predicted pulmonary defect agreed with a measured â?¶â?¸Ga-PET pulmonary defect.
Results: Training accuracy saturated at 6 fully-connected hidden layers (Ï?=0.98, r²=0.97, RMSE=31), where each model best approximated its ground-truth â?¶â?¸Ga-PET counterpart. Intra-patient and inter-patient model performance statistics were: (Ï?=0.91,r²=0.89,RMSE=48) and (Ï?=0.64,r²=0.65,RMSE=107), respectively. Intra-patient training, intra-patient testing, and inter-patient validation achieved DSC values of 0.91, 0.79, and 0.74, respectively. This indicates reasonably high model generalization, and the ability to potentially identify defects in pulmonary ventilation.
Conclusion: By combining quantitative radiomic filtering with deep CNNs, this study demonstrated a novel methodology to decode pulmonary function embedded within CT images.
Funding Support, Disclosures, and Conflict of Interest: This research is partially supported by a grant from Varian Medical Systems.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
