
Contact the VUNO team anytime.
RSNA 2018
Purpose To evaluate the effectiveness of a slice thickness reduction technique in computed tomography(CT) scans using convolutional neural network(CNN)-based super-resolution(SR) network for improving the sensitivity of lung nodule detection in thick section CT scans.
Method and Materials We collected 100 sets of CT scans with identical acquisition protocols that were differentiated only by the slice thickness (1mm, 3mm, and 5mm). By employing CNN-based SR network, we trained the model to learn the residuals between synthesized thin section slices and real thin section slices. We used 80 sets of CT scans for training, 5 sets for the validation and the remaining 10 sets were used for quantitative evaluation. We separately collected 100 sets of CT scans (also with 1mm, 3mm, and 5mm slice thickness) with one biopsy-confirmed nodule per scan(46 solid nodules and 54 non-solid nodules, size ranges 6-12mm) to evaluate the effectiveness of the slice thickness reduction techniques for improving the lung nodule detection performance in thick section CT scans. The computer-aided detection(CAD) system used for the evaluation of lung nodule detection performance was internally developed with LUNA16 dataset, which contains 888 CT scans with slice thickness less than 3mm.
Results When slice thickness was reduced from 3mm to 1mm, the peak signal-to-noise ratio(PSNR) and structural similarity index(SSIM) were 30.9624 and 0.8142 respectively; when slice thickness was reduced from 5mm to 1mm PSNR and SSIM were 29.2620 and 0.7439 respectively. In the nodule detection task, all solid and non-solid nodules were detected by the CAD system using 1mm slice thickness scans(100.0% recall). However, 2 solid nodules were missed when using 3mm slice thickness scans (95.7% recall) while their corresponding synthetic 1mm scans improved the recall to 97.8% (1 missed solid nodule). Recall of scans with 5mm slice thickness was 89.1% and 85.2% while their synthetic 1mm scans improved the recall to 100.0% and 96.3% for solid and non-solid nodules respectively.
Conclusion Our CNN-based SR method generates synthetic thin section slices from thick section slices which improve lung nodule detection performance using CAD.
Clinical Relevance/Application Robustness to acquisition protocol is essential for reliable lung CAD systems. Our slice thickness reduction technique may improve the robustness of CAD systems when applied to CT scans with various slice thickness.