SS 604 - Pulmonary nodules and screening
Subsolid and part-solid nodules in lung cancer screening: comparison between visual and computer-aided detection
Purpose: To compare visual detection and computer aided diagnosis (CAD) for detection of non-solid nodules (NSN) and part-solid nodules (PSN), in a lung cancer screening trial.
Methods and Materials: Baseline low-dose computed tomography (LDCT) of 2303 subjects were assessed by 2 independent operators: a) visual detection (VD); b) CAD software (CIRRUS Lung Screening). CAD was run also on first and second incidence round. Rate of agreement was calculated by weighted k test. Sensitivity and negative predictive value (NPV) were calculated according to cumulative number of subjects with detected nodules (VD and CAD). LDCT features were compared between CAD-only and VD-only detected nodules.
Results: Nodules were detected in 215/2303 subjects (dominant nodule: 171/215 NSN and 44/215 PSN), notably 149 were CAD-only detected (113 NSN and 36 PSN), 27 were VD-only detected (25 NSN and 2 PSN), and 39 were CAD and VD detected (33 NSN and 6 PSN). The agreement was fair (weighted k=0.276). Sensitivity and NPV for CAD 87.4% and 98.7%, and for VD 30.7% and 93.3%. Automatic and manual caliper were similar for the assessment of maximum diameter (p=0.111). CAD-only and VD-only detected nodules showed similar nodule diameter (p=0.727) and proportion of PSN (p=0.073). Volume of CAD-only detected nodules was significantly smaller (p=0.019). Among the 27 VD-only detected nodules, 11 NSN were subsequently detected by CAD at the incidence rounds.
Conclusion: Detection of NSN and PSN showed only fair agreement between VD and CAD. Combined assessment of LDCT by both VD and CAD is needed to achieve optimal sensitivity.
'Non-nodule' appearance of early lung cancer in CT screening: retrospective evaluation of 281 lung cancers detected
Purpose: Pulmonary nodule is the typical presentation of lung cancer. However, there is a spectrum of focal lesions that cannot be defined as pulmonary nodule, but really represent early manifestation of lung cancer. Therefore, we aimed to evaluate the features of early stage lung cancers which are depicted on low-dose CT as 'non-nodule' lesion.
Methods and Materials: Two experienced radiologists in lung CT screening retrospectively reviewed the radiological features of 281 lung cancers detected during 10 years of screening trial (5203 heavy smokers, >50-year-old). The readers in consensus assessed if lung cancer appearance at time of diagnosis was consistent (‘nodule’) or not (‘non-nodule’) with the Fleischner Society definition of ‘pulmonary nodule’. ‘Non-nodule’ lung cancers were further classified as: 1) bulla-like, 2) scar-like, 3) endobronchial, 4) other. In case of disagreement, a third radiologist (>15-year experience) resolved the discrepancies.
Results: Twenty-four lung cancers (24/281, 8.5%) were defined as ‘non-nodule’ lesions: 7 (2.9%) were bulla-like, 9 (3.7%) scar-like, 4 (1.6%) endobronchial and 4 (1.6%) other. The mean diameter at time of diagnosis was 16.2 mm (range: 7.5-65 mm). Nineteen were adenocarcinomas, two SCC, two NSCLS non-otherwise-specified, and one SCLC. Nineteen were diagnosed in stage I, four in stage III and one in stage IV.
Conclusion: A non-negligible proportion (8.5%) of screening-detected lung cancers has a ‘non-nodule’ appearance. Bulla-like and scar-like lesions are predominant and should be carefully evaluated. The awareness of these atypical lung cancer presentations can avoid missed diagnosis and should be considered in management of screening-detected focal lesions.
The implications of internal vessel and bronchial changes within pure ground-glass opacity lung adenocarcinoma on CT
Purpose: To investigate the implications of vessel and bronchial changes within lesion on CT by comparing with histopathologic subtypes of lung adenocarcinoma with pure ground-glass nodule (pGGN).
Methods and Materials: One hundred and ninety patients (201 lesions) of lung adenocarcinomas with pGGN who underwent curative resection were included. All patients were categorized into two groups according to the abnormalities of vessel and bronchus: abnormal group and normal group. Histopathologic subtypes were classified into preinvasive lesions (atypical adenomatous hyperplasia [AAH] and adenocarcinoma in situ [AIS]), minimally invasive adenocarcinoma (MIA) and invasive pulmonary adenocarcinoma (IPA). Pearson Chi square test was used to analyze the relationships between vessel and bronchial abnormalities and histopathologic subtypes. Mann-Whitney rank test and t-test were used to identify the correlation of vessel and bronchial abnormalities with the density and diameter of pGGN.
Results: There were 4 with vessel abnormalities and 1 with bronchial abnormalities in 35 preinvasive lesions, 26 and 18 in 53 MIAs, 80 and 55 in 113 IPAs, respectively. There were statistically significant differences of vessel and bronchial abnormalities among histopathologic subtypes (P=0.000, 0.000). There were significant differences in vessel and bronchial abnormalities with pGGN diameter (P=0.000, 0.000), but not with density (P=0.303, 0.966). There were no significant difference between vessel convergence and vessel dilation or distortion among these three subtypes (P=0.115).
Conclusion: Vessel and bronchial abnormalities within the pGGN may indicate the invasiveness of lung adenocarcinoma and were correlated with the lesion diameter.
Quantitative analysis for determining the optimal computed tomography threshold value to detect invasive foci in subsolid nodules
Purpose: To establish the optimal CT threshold value for detecting the solid components of subsolid nodules, using invasive foci as reference.
Methods and Materials: Thin-section non-enhanced chest CT scans were retrospectively reviewed for 25 patients (M:F=7:18; age, 55.9±12.7 years) with pathologically confirmed MIA. The solid portion was assumed as the surrogate marker of invasive foci of MIA. It was defined as the area with higher attenuation than the threshold and measured using the maximum diameter on multiplanar reconstruction images with various HU thresholds ranging from −600 to −100 HU in 50-HU intervals. A linear mixed model for considering repeated measures from various thresholds was used to evaluate bias in each threshold, using the pathological size of invasive foci as the reference.
Results: The size of the 25 MIA was 9.2±1.33 mm, with mean attenuation of −539.1±96 HU. The size of the pathologically invasive foci was 3.56±1.33 mm. The maximum size (mm) of the solid portion was 0.31±0.87 on −100 HU, 0.55±1.06 on −150 HU, 0.93±1.31 on −200 HU, 1.39±1.6 on −250 HU, 2.00±1.67 on −300 HU, 2.76±2.22 on −350 HU, 3.58±2.39 on −400 HU, 4.52±2.39 on −450 HU, 5.44±2.47 on −500 HU, 6.67±2.34 on −550 HU, and 7.75±2.56 on −600 HU. At the threshold of −400 HU, the bias was lowest (−0.0176) between the solid portion and invasive foci, showing an insignificant difference (p=0.963).
Conclusion: For quantitative analysis, −400 HU might be the optimal threshold to define the solid portion of subsolid nodules as a surrogate marker of invasive foci.
Comparison of lung nodule detection performance on 3D CAD system among filtered back projection and iterative reconstruction methods at different radiation-dose levels
Purpose: To compare the nodule detection capability on a 3D computer-aided detection (CAD) system among filtered-back projection (FBP) and iterative reconstruction (IR) methods at standard-, reduced- and ultra-low-dose CTs (SDCT, RDCT and ULDCT).
Methods and Materials: Forty patients prospectively underwent chest CT examinations with SDCT (250mA), RDCT (50mA) and ULDCT (10mA) protocols, and CT data were reconstructed into 1-mm-thick images with and without commercially available IR method (i.e. AIDR 3D). Therefore, the following CT data set in each patient was reconstructed: SDCT with and without AIDR 3D, RDCT with and without AIDR 3D, and ULDCT with and without AIDR 3D. Then, nodule detections were automatically performed by our proprietary CAD software. To determine the utility of IR method for improving nodule detection capability, sensitivity and false positive rate (/case) of the CAD system were also compared among all protocols by means of McNemar’s test or signed rank test.
Results: Although there were no significant difference of false-positive rate among all protocols, sensitivities of RDCT and ULDCT with AIDR 3D (RDCT: 72.3%, ULDCT: 66.3%) were significantly higher than that without AIDR 3D (RDCT: 56.4%, p<0.0001; ULDCT: 35.6%, p<0.0001). Sensitivity of SDCT with and without AIDR 3D (with AIDR 3D: 73.3%, without AIDR 3D: 76.2%) were significantly higher than that of RDCT without AIDR 3D (p<0.0001) and ULDCT with and without AIDR 3D (p<0.0001).
Conclusion: Iterative reconstruction method is useful for improving nodule detection performance on a 3D CAD system at reduced- and ultra-low-dose CTs as compared with filtered-back projection method.
Effect of detectability of pulmonary nodules with lowering dose based on nodule size, type and body mass index with different iterative reconstruction algorithms
Purpose: To determine the diagnostic accuracy of lung nodule detection in thoracic CT using 2 reduced-dose protocols comparing 3 available CT reconstruction algorithms (filtered back projection-FBP, adaptive statistical reconstruction-ASIR and model-based iterative reconstruction-MBIR) in a western population.
Methods and Materials: A prospective single-center IRB-approved study recruited 98 patients with written consent. Standard-dose (STD) thoracic CT followed by 2 reduced-dose protocols using automatic tube current modulation (RD1) and fixed tube current (RD2) were performed and reconstructed with FBP, ASIR and MBIR with subsequent diagnostic accuracy analysis for nodule detection.
Results: 108 solid nodules, 47 subsolid nodules and 89 purely calcified nodules were analysed. RD1 was superior to RD2 for assessment of solid nodules ≤4mm, and subsolid nodules ≤5mm. Deterioration of RD2 is correlated to patient’s body mass index and least affected by MBIR. For solid nodules ≤4mm, MBIR area under curve (AUC) for RD1 was 0.935/0.913 and AUC for RD2 was 0.739/0.739, for rater 1/rater2 respectively. For subsolid nodules ≤5mm, MBIR AUC for RD1 was 0.971/0.986 and AUC for RD2 was 0.914/0.914, for rater 1/rater2 respectively. For calcified nodules excellent detection accuracy was maintained regardless of reconstruction algorithms with AUC >0.97 for both readers across all dose and reconstruction algorithms.
Conclusion: Diagnostic performance of lung nodule is affected by nodule size, protocol, reconstruction algorithm and patient’s body habitus. The protocol in this study showed that RD1 was superior to RD2 for assessment of solid nodules ≤4mm, and subsolid nodules ≤5mm and deterioration of RD2 is related to patient’s body mass index.
Influence of reconstruction methods to measurement accuracy for computer-aided volumetry (CADv) at standard-, reduced-, low-dose and ultra-low-dose CT in QIBA phantom study
Purpose: To directly compare the capability of three reconstruction methods using forward projected model-based iterative reconstruction (FIRST), adaptive iterative dose reduction using 3D processing (AIDR 3D) and filtered back projection (FBP) for radiation dose reduction and accuracy of computer-aided volumetry (CADv) measurements on chest CT examination in the QIBA recommended phantom study.
Methods and Materials: An thoracic phantom with 30 simulated nodules of three density types (100, -630, and -800 HU) and five different diameters was scanned with an area-detector CT at tube currents of 270, 80, 40, 20, and 10 mA. Each scanned data was reconstructed with three methods. CT value and image noise were measured, and compared among three reconstruction methods at each tube current by Tukey’s HSD test. For comparison of the capability for CADv at each tube current, Tukey’s HSD test was used to compare the percentage of absolute measurement errors for three reconstruction methods.
Results: Image noises of FIRST and AIDR 3D for each nodule type were significantly lower than that of FBP at each tube current (p<0.05). In addition, image noises of FIRST were significantly lower than that of AIDR 3D at all tube currents except 270mA (p<0.05). Mean absolute measurement errors of AIDR 3D and FIRST for each nodule type were significantly lower than those of the FBP method at 20mA and 10mA (p<0.05).
Conclusion: FIRST and AIDR 3D methods are more effective than the FBP method for radiation dose reduction, while yielding better measurement accuracy of CADv for chest CT examination.
Effect of image reconstruction with ADMIRE in low-dose and ultra-low-dose CT volumetry of solid pulmonary nodules: a phantom study
Purpose: To investigate the influence of ultra-low-dose CT and iterative reconstruction with ADMIRE on volumetry of solid lung nodules.
Methods and Materials: CT scans were obtained with third-generation dual-source CT at standard dose, 1/8, 1/20 and 1/70 of standard dose using an anthropomorphic chest phantom containing solid microspheres (4-10mm). An extension ring was used for different phantom sizes (M, L). Image reconstruction was performed with FBP, ADMIRE3 and 5. Three oberservers independently measured nodule volumes semi-automatically. Relative percentage error (RPE) was computed to evaluate measurement accuracy. To assess influence of various parameters on RPE univariate ANOVA with post hoc Bonferroni-Test was applied. P<.05 was considered significant.
Results: Intra-class correlation coefficient for RPE between observers was 0.862 (95%CI: 0.838-0.883). Phantom size, nodule diameter, CT dose protocol and reconstruction algorithm significantly contributed to RPE (all p<.001). Underestimation of mean nodule volume among all dose protocols was -5.3±6.0% with ADMIRE3,-7.0±5.6% with ADMIRE5, and-2.7±8.5% with FBP (all p<.001). Overall scanning with 1/70 dose protocol showed higher underestimation of mean volume (-7.3±10.3%) compared with standard dose (-4.1±4.0%), 1/8 (-4.6±4.3%) and 1/20 (-4.1±7.3%) dose (all p<.001). There were no significant differences between other dose values. Overall volume underestimation was highest in 6mm nodules (-10.3±5.4%), compared to 4 (-1.3±10.3%), 8 (-3.2±2.8%) and 10mm (-4.7±1.9%) nodules. Phantom size L (-5.8±7.9%) showed significant lower overall mean RPE compared to size M (-4.2±6.0%; p<.001).
Conclusion: Our study indicates that image reconstruction with ADMIRE3 and 5 and scanning with 1/70 dose leads to little but significant underestimation of solid nodule volume.
Comparison of the effect of model-based iterative reconstruction (MBIR) and filtered back projection (FBP) algorithms on software measurements in pulmonary ground-glass nodules
Purpose: To evaluate the differences between filtered-back-projection (FBP) and model-based iterative reconstruction (MBIR) algorithms on semi-automatic measurements in subsolid nodules (SSNs).
Methods and Materials: Unenhanced CT-scans of 73 SSNs obtained using the same protocol and reconstructed with both FBP and MBIR algorithms were evaluated by two radiologists. Diameter, mean-attenuation, mass, and volume of whole nodules and its solid components were measured. Intra-, inter-observer variability and differences between FBP and MBIR were then evaluated using Bland-Altman method and Wilcoxon tests.
Results: Longest diameter, volume and mass of nodules and those of its solid component were significantly higher using MBIR (p<0.05) with mean differences of 1.1% (limits of agreement, -6.4 to 8.5%), 3.2% (-20.5 to 27%), 3.2% (-20.9 to 27.3%), 6.3% (-51.9 to 64.6%), 2.9% (-16.9 to 22.7%), and 6,6% (-50.1 to 63.3%), respectively. The limits of agreement between FBP and MBIR were within the range of intra- and inter-observer variability for both algorithms with respect to the diameter, volume and mass of nodules and its solid components. There were no significant differences in intra- or inter-observer variability between FBP and MBIR (p>0.05).
Conclusion: Semi-automatic measurements of SSNs significantly differed between FBP and MBIR, however, the differences were within the range of measurement variability.
Purpose: One of the most performed examinations by radiologists is chest radiography, providing information of important anatomical structures of the body. Reporting chest x-rays is a demanding task and very important medical-legally, sometimes neglected. The aim is setting up a screening tool for giving priority to the abnormal radiographs and facilitate the reporting task. New techniques based on artificial intelligence have emerged (field of computer vision). We propose to use deep learning (convolutional neural networks, CNN) in a computer-aided diagnostic (CAD) system to help radiologists to perform automated screening of chest radiographs.
Methods and Materials: A database of chest x-ray images (7470) from the Indiana University was used for the study. A subset of cases was selected consisting of 2242 A-P, divided into different groups (abnormal, cardiomegaly, nodule, opacity, atelectasis, pleural effusion, each one versus normal). The AlexNet CNN (pretrained with ImageNet) was implemented to extract features from the different groups needed to train binary support vector machine classifiers. An NVIDIA Tesla K40 GPU was used to optimize the computing performance.
Results: A software tool was successfully implemented obtaining an area under curve (AUC) of 0.86 for abnormal, 0.91 for cardiomegaly, 0.92 for pleural effusion, 0.70 for nodule, 0.90 for atelectasis group and 0.86 for opacity.
Conclusion: Using a pretrained CNN as an automatic feature extractor for training an SVM classifier is a good approach to get relevant results in chest x-ray screening tasks.