(Tuesday, 7/16/2019) 9:30 AM - 10:00 AM
Room: Exhibit Hall | Forum 6
Purpose: Most Computer Aided Diagnosis (CAD) tools have been trained on cases from conventional (diagnostic) dose exams. However, the robustness of CAD performance in low-dose (screening) conditions with iterative reconstruction algorithms has not been evaluated. The purpose of this study was to investigate the effects of these conditions on CAD in lung nodule detection within the low-dose screening paradigm.
Methods: Raw data of low-dose, chest CT images were acquired from 98 screening patients (58 solid nodules, 9 part-solid, Table 1). Dose reduction was simulated via a validated noise-addition model, simulating reduced doses to 50%, 25%, and 10% of original dose (1-2mGy). Images were reconstructed at a scanner (Siemens Definition AS; Siemens Healthineers, Erlangan, Germany) at different combinations of iterative SAFIRE kernels and slice thicknesses (Table 2)– including 3 common parameter sets, 1 extreme smooth (ExSm), and 1 extreme sharp (ExSh). Lung nodule CAD generated ROIs of suspected lesions for each case. The ROIs were compared to reference truth generated from physician reports.
Results: Mean sensitivity for detection of solid nodules seemed higher on conditions with smoother kernels and larger slice thickness (Fig. 1, top). ExSm maintained sensitivity performance through changes in dose, while ExSh had worse performance across all doses, with higher sensitivity to dose reduction. CAD had difficulty detecting part-solid nodules, with only three conditions above 45% mean sensitivity. Median false positives rates were stable across all conditions, save sharp kernels below 50% dose.
Conclusion: Iterative SAFIRE reconstructions are more sensitive to dose reduction at sharper kernels and smaller slice thicknesses. Sensitivity of CAD (initially tuned on solid nodules in diagnostic dose conditions) was below average, as the dataset comprised of conditions beyond the already low-dose acquisitions. To address this, future work includes noise mitigation/image homogenization strategies, or added false positive reduction modules to enable raised detection sensitivity.
Keywords
Low-dose CT, CAD, Computer Vision
Taxonomy
Contact Email
