Room: Exhibit Hall | Forum 6
Purpose: To assess the optimal operation ranges of CT image acquisition and reconstruction parameters for automatic contouring algorithms.
Methods: Eight head-and-neck patientsâ€™ CT scans were reconstructed from the sinograms with varying slice thicknesses (0.6-10mm) and pixel sizes (0.39-1.17mm). CT dose was studied by adding noise prior to reconstruction using a Siemens low-dose simulation software (simulated mAs range: 10%-100% of original). The impact of these imaging parameters on two in-house auto-contouring algorithms, one convolutional neural network(CNN)-based and one multi-atlas-based, was investigated on a total of 133 reconstructed scans. Normal tissue auto-contours were compared with the results from scans with 3mm slice thickness, 0.977mm pixel size, and the 100% CT dose.
Results: Slice thicknesses between 2-4mm resulted in an average Dice and Hausdorff distance (HD) of 0.76 and 0.37cm. The average Dice was 0.82 and 0.70 for the CNN-based and the multi-atlas-based algorithms, respectively. Outside of this range, contour agreement worsened for both algorithms (average Dice and HD of 0.70 and 0.52cm). Pixel sizes between 0.78-1.17mm resulted in an average Dice and HD of 0.89 and 0.29cm. The average Dice was 0.90 and 0.87 for the CNN-based and the multi-atlas-based algorithms, respectively. Outside of this range, contour agreement worsened for both algorithms (average Dice and HD of 0.65 and 0.51cm). Even when the image was simulated with 10% of its original dose, the average Dice for the CNN-based and the multi-atlas-based algorithms were 0.96 and 0.74 and the average HD were 0.17cm and 0.58cm respectively.
Conclusion: We found a range of pixel size and slice thickness for which algorithm contours were consistent. Additionally, we found that the two auto-contouring algorithms were relatively insensitive to changes in CT dose. Overall, the CNN-based algorithm was less sensitive to changes in the three imaging parameters than the multi-atlas-based algorithm.