Usefulness of dictionary learning-based processing for improving image quality of submillisievert low-dose chest CT: initial experience

概要

Purpose:
　The purposes of this study were to develop a dictionary learning (DL)-based post-processing technique for low-dose chest computed tomography (CT), and to evaluate the usefulness of DL- based post-processing for improving image quality on sub-millisievert chest CT.

Materials and methods:
　Standard-dose and sub-millisievert chest CT were acquired in 12 patients. Dictionaries including standard- and lowdose image patches were generated from the CT datasets. For each patient, DL-based processing was performed for low-dose CT using the dictionaries generated from the remaining 11 patients. This procedure was repeated for all12 patients. Objective assessment of image noise was performed to evaluate the image qualities of sub­ millisievert CT with and without DL-based processing. For this assessment, the standard deviation of CT density in the descending aorta at the level of the tracheal carina was measured by placing regions of interest within the descending aorta. Image quality of normal thoracic structures at the 3 levels of the upper mediastinum, azygos arch and left atrium on the sub- millisievert chest CT images was visually assessed on a 5-point scale (5二excellent,l=very poor). Lesion conspicuity of lung nodules (solid nodules, n二14, mean lesion size 9.3 ± 6.9 mm; ground-glass nodules, n二6, mean lesion size 9.4 ± 4.2 mm) in the subjects was visually assessed on a 5-point scale as well.

Results:
　Image noise on sub-millisievert CT was significantly decreased with DL'based image processing (48.5±13.7 HU vs 20.4±7.9 HU, p=0.0005). Image quality scores for normal mediastinal structures were significantly improved by DL-based processing (upper level' 1.58 ± 0.51 vs 2.67 ± 0.49, p=0.001; middle level：1.92 ± 0.67 vs 2.83 ± 0.58, p=0.0039; lower level:1.58 ± 0.51 vs 2.92 ± 0.67, p=0.001). As for image quality scores for normal lung structures, we found a significant improvement after DL processing at the middle and lower levels of the chest (middle level- 2.25 ± 0.75 vs 2.92 ± 0.79, p=0.0078； lower level- 2.25 ± 0.87 vs 3.50 ± 0.67, p=0.0078). Lung lesion conspicuity on sub-millisievert chest CT was also significantly improved by DL-based processing (solid nodules- 2.7 ± 0.6 vs 3.4 ± 0.6, p = 0.027； ground-glass nodules, 2.8 ± 0.4 vs 4.2 ± 0.8, p = 0.031).

Conclusion:
　Image quality on sub-millisievert chest CT images can be significantly improved by DL-based post-processing utilizing dictionary pairs of standard- and low-dose chest CT. Using this DL-based technique, image noise was reduced by approximately 60% on sub-millisievert chest CT. We also found that the DL-based post-processing technique substantially improved the diagnostic assessability of thoracic structures on sub-millisievert chest CT. The proposed technique may facilitate the use of sub-millisievert CT for lung cancer screening.