Augmenting motion artifacts to enhance auto-contouring of complex structures in cone-beam computed tomography imaging.

Abstract

Objective. To develop an augmentation method that simulates cone-beam computed tomography (CBCT) related motion artifacts, which can be used to generate training-data to increase the performance of artificial intelligence models dedicated to auto-contouring tasks.Approach.The augmentation technique generates data that simulates artifacts typically present in CBCT imaging. The simulated pseudo-CBCT (pCBCT) is created using interleaved sequences of simulated breath-hold and free-breathing projections. Neural networks for auto-contouring of head and neck and bowel structures were trained with and without pCBCT data. Quantitative and qualitative assessment was done in two independent test sets containing CT and real CBCT data focus on four anatomical regions: head, neck, abdomen, and pelvis. Qualitative analyses were conducted by five clinical experts from three different healthcare institutions.Main results.The generated pCBCT images demonstrate realistic motion artifacts comparable to those observed in real CBCT data. Training a neural network with CT and pCBCT data improved Dice similarity coefficient (DSC) and average contour distance (ACD) results on CBCT test sets. The results were statistically significant (p-value ⩽.03) for bone-mandible (model without/with pCBCT: 0.91/0.92 DSC,p⩽ .01; 0.74/0.66 mm ACD,p⩽.01), brain (0.34/0.93 DSC,p⩽ 1 × 10^-5; 17.5/2.79 mm ACD,p= 1 × 10^-5), oral-cavity (0.81/0.83 DSC,p⩽.01; 5.11/4.61 mm ACD,p= .02), left-submandibular-gland (0.58/0.77 DSC,p⩽.001; 3.24/2.12 mm ACD,p⩽ .001), right-submandibular-gland (0.00/0.75 DSC,p⩽.1 × 10^-5; 17.5/2.26 mm ACD,p⩽ 1 × 10^-5), left-parotid (0.68/0.78 DSC,p⩽ .001; 3.34/2.58 mm ACD,p⩽.01), large-bowel (0.60/0.75 DSC,p⩽ .01; 6.14/4.56 mm ACD,p= .03) and small-bowel (3.08/2.65 mm ACD,p= .03). Visual evaluation showed fewer false positives, false negatives, and misclassifications in artifact-affected areas. Qualitative analyses demonstrated that, auto-generated contours are clinically acceptable in over 90% of cases for most structures, with only a few requiring adjustments.Significance.The inclusion of pCBCT improves the performance of trainable auto-contouring approaches, particularly in cases where the images are prone to severe artifacts.