Room: Davidson Ballroom B
Purpose: Low-dose CT has been extensively used in lung cancer screening. However, lowering radiation dose results in more noise and thus impedes subsequent diagnosis. The purpose of this work was to investigate the performance of a Convolutional Neural Network (CNN) for lung segmentation, trained using CT scans acquired at diagnostic dose, when applied to low and ultra-low dose scans.
Methods: Clinical data and corresponding raw projection data of 40 different patients were acquired, comprised of a cohort of 30 from lung cancer screening (~2 mGy) and the other 10 from diffuse lung disease trials (~15 mGy). Further reduced dose scans at 50%, 25%, and 10% of the original dose were simulated by adding calibrated noise to the projection data and they were all reconstructed with 1mm slice thicknesses/spacing and medium kernel. A previous ResNet-101 based CNN lung segmentation model built only using diagnostic scans was applied to segment these reconstructed scans of different dose levels. The ground truth segmentation was derived from the clinically reconstructed image by a semi-automated lung segmentation method, followed by manual editing. Dice Similarity Coefficient (DSC) and Average Surface Distance (ASD) were used as evaluation metrics.
Results: For the lung cancer screening cohort (30 patients), CNN achieved a mean DSC of 0.987, 0.985, 0.982 and 0.978 at 100%, 50%, 25% and 10% CT doses, respectively. The corresponding ASDs were 0.826 mm, 0.917 mm, 1.059 mm and 1.218 mm. For the diffuse lung disease cohort (10 patients), the CNN achieved 0.978, 0.973, 0.969 and 0.961 at 100%, 50%, 25% and 10% doses, respectively. The corresponding ASDs were 0.842 mm, 0.926 mm, 1.083 mm and 1.242 mm.
Conclusion: The CNN based lung segmentation model trained on diagnostic dose CT scans (~ 15 mGy) was able to achieve accurate segmentation on simulated low and ultra-low dose scans (~0.2 mGy).
Low-dose CT, CAD, Segmentation