Balancing Prevalence of Contrast and Non-Contrast Computed Tomography Examples in a Limited Set and Training Transformer-Based Great Vessel Segmentation.

Sparing normal cardiac and vascular substructures from radiation is important to minimize toxicity. Hence, deep learning transformer models were implemented with emphasis on prevalence and number of contrast-enhanced (CECT) and non-contrast CT (NCCT) images in the training set for segmenting the great vessels. Two hundred and forty CTs of patients diagnosed with locally advanced non-small cell lung cancer (LA-NSCLC) were evaluated. A 3D transformer encoder with U-Net decoder was fine-tuned to segment five great vessels: the aorta, pulmonary artery (PA), pulmonary vein (PV), superior vena cava (SVC), and inferior vena cava (IVC) using 3-fold cross-validation. The transformer encoder was previously trained with self-supervised learning on a large number (n = 10,000) of unlabeled CT scans sourced from multiple disease sites using institutional and public datasets. Three model instances with distinct dataset configurations were analyzed: (a) a model trained on the entire training set with (n = 60) CECT and (n = 120) NCCT scans, (b) a model trained only on the CECT subset of (n = 60) scans, and (c) a model trained on the balanced CECT (n = 30) and NCCT (n = 30) scans. Testing was done on 60 patients not used for training. Segmentation accuracy was measured using the Dice Similarity Coefficient (DSC), and the statistical differences between the methods were analyzed using a two-sided, paired, Wilcoxon signed-rank test with Bonferroni correction used for multiple comparisons. Summary results on the testing set consisting of 25 CECTs and 35 NCCT images is shown in Table 1. Balanced and the entire training set models were not significantly different for the analyzed vessels overall as well as for contrast and non-contrast scans (p > 0.05). The CECT model was similarly accurate as the entire training set model for CECT for all vessels and aorta, SVC and IVC for NCCT. The CECT model had significantly lower accuracies for PA (p = 0.015) and PV (p = 0.010) on the NCCT scans than the entire training set model. Our results showed that a model trained with smaller but balanced prevalence of CECT and NCCT scans was similarly accurate as a larger (entire) training set model. In addition, the CECT only model was similarly accurate as the entire training set model for CECT but less so for two of the five analyzed vessels. These results indicate the possibility of using fewer but easily delineated cases with less uncertainty in manually identifying such structures to train generalizable AI methods. [ABSTRACT FROM AUTHOR]