"KAIZEN" method realizing implementation of deep-learning models for COVID-19 CT diagnosis in real world hospitals.

Numerous COVID-19 diagnostic imaging Artificial Intelligence (AI) studies exist. However, none of their models were of potential clinical use, primarily owing to methodological defects and the lack of implementation considerations for inference. In this study, all development processes of the deep-learning models are performed based on strict criteria of the "KAIZEN checklist", which is proposed based on previous AI development guidelines to overcome the deficiencies mentioned above. We develop and evaluate two binary-classification deep-learning models to triage COVID-19: a slice model examining a Computed Tomography (CT) slice to find COVID-19 lesions; a series model examining a series of CT images to find an infected patient. We collected 2,400,200 CT slices from twelve emergency centers in Japan. Area Under Curve (AUC) and accuracy were calculated for classification performance. The inference time of the system that includes these two models were measured. For validation data, the slice and series models recognized COVID-19 with AUCs and accuracies of 0.989 and 0.982, 95.9% and 93.0% respectively. For test data, the models' AUCs and accuracies were 0.958 and 0.953, 90.0% and 91.4% respectively. The average inference time per case was 2.83 s. Our deep-learning system realizes accuracy and inference speed high enough for practical use. The systems have already been implemented in four hospitals and eight are under progression. We released an application software and implementation code for free in a highly usable state to allow its use in Japan and globally.

The effect of the virtual monochromatic spectral imaging for the metallic artifact and the pulmonary nodule detection.

OBJECTIVES: This study aimed to evaluate whether dual-energy computed tomography can reduce metal artifacts and improve detection of pulmonary nodules. METHODS: Twelve simulated nodules were randomly placed inside a chest phantom with a pacemaker. Then, dual-energy computed tomography was performed, and 5 virtual monochromatic images at 40, 50, 65, 100, and 140 keV were reconstructed with 5- and 0.625-mm slice thicknesses. Two independent observers assessed the metal artifact (3-point scale from 1, none, to 3, severe) and detection of the nodule (5-point scale from 1, definitely absent, to 5, definitely present). Statistical analysis was performed with a P value of less than 0.01 (0.05/5). RESULTS: With both slice thicknesses, the metallic artifact increased at 40 or 50 keV and decreased at 100 or 140 keV relative to that at 65 keV (P < 0.01). The nodule detection score was not significantly different between each kiloelectron volt level with the 0.625-mm slice thickness; however, the score was significantly worse at 40 keV compared to 65 keV (P < 0.01) with the 5-mm slice thickness. CONCLUSIONS: High monochromatic energy images can reduce metal artifacts without a change in nodule detection score. Low monochromatic energy images increase metal artifacts and worsen nodule detection in thick slices.

Lung image quality with 320-row wide-volume CT scans: the effect of prospective ECG-gating and comparisons with 64-row helical CT scans.

RATIONALES AND OBJECTIVES: To evaluate the image quality of 320-row wide-volume (WV) computed tomography (CT) scans in comparison with 64-row helical scans for the lung. MATERIALS AND METHODS: The Institutional Review Board of each institution approved this prospective, multicenter study and informed consent was obtained. A total of 73 subjects underwent two types of chest CT, including 320-row WV scans and 64-row helical scans. Both scans used the same tube voltage, tube current, exposure time setting, and slice thickness. The helical scans were not electrocardiogram (ECG)-gated. For the WV scans, prospective ECG-gating was used for 38 subjects, whereas the other 35 subjects did not have ECG-gating. Using a 5-point scale from 1 (nondiagnostic) to 5 (excellent), three blinded observers independently evaluated image quality for five lobes and the lingula. The differences in the scores between WV scans and helical scans were compared using Wilcoxon's signed-rank test. RESULTS: The WV scans with ECG-gating had significantly higher scores than 64-row helical scans for all lobes and lingula (right lower lobe, P < .01; other lobes and lingula, P < .0001, respectively). The 320-row WV scans without ECG-gating also had significantly higher scores than 64-row helical scans (P < .05), except for nonsignificant differences for the left upper lobe. CONCLUSIONS: Lung image quality of ECG-gated WV scans, which do not require any additional radiation exposure, is better than that of non-ECG-gated 64-row helical scans. Non-ECG-gated 320-row WV scans are comparable or slightly superior to non-ECG-gated 64-row helical scans.

Pulmonary nodules: effect of adaptive statistical iterative reconstruction (ASIR) technique on performance of a computer-aided detection (CAD) system-comparison of performance between different-dose CT scans.

PURPOSE: To evaluate the effects of ASIR on CAD system of pulmonary nodules using clinical routine-dose CT and lower-dose CT. MATERIALS AND METHODS: Thirty-five patients (body mass index, 22.17 ± 4.37 kg/m(2)) were scanned by multidetector-row CT with tube currents (clinical routine-dose CT, automatically adjusted mA; lower-dose CT, 10 mA) and X-ray voltage (120 kVp). Each 0.625-mm-thick image was reconstructed at 0%-, 50%-, and 100%-ASIR: 0%-ASIR is reconstructed using only the filtered back-projection algorithm (FBP), while 100%-ASIR is reconstructed using the maximum ASIR and 50%-ASIR implies a blending of 50% FBP and ASIR. CAD output was compared retrospectively with the results of the reference standard which was established using a consensus panel of three radiologists. Data were analyzed using Bonferroni/Dunn's method. Radiation dose was calculated by multiplying dose-length product by conversion coefficient of 0.021. RESULTS: The consensus panel found 265 non-calcified nodules ≤ 30 mm (ground-glass opacity [GGO], 103; part-solid, 34; and solid, 128). CAD sensitivity was significantly higher at 100%-ASIR [clinical routine-dose CT, 71% (overall), 49% (GGO); lower-dose CT, 52% (overall), 67% (solid)] than at 0%-ASIR [clinical routine-dose CT, 54% (overall), 25% (GGO); lower-dose CT, 36% (overall), 50% (solid)] (p<0.001). Mean number of false-positive findings per examination was significantly higher at 100%-ASIR (clinical routine-dose CT, 8.5; lower-dose CT, 6.2) than at 0%-ASIR (clinical routine-dose CT, 4.6; lower-dose CT, 3.5; p<0.001). Effective doses were 10.77 ± 3.41 mSv in clinical routine-dose CT and 2.67 ± 0.17 mSv in lower-dose CT. CONCLUSION: CAD sensitivity at 100%-ASIR on lower-dose CT is almost equal to that at 0%-ASIR on clinical routine-dose CT. ASIR can increase CAD sensitivity despite increased false-positive findings.

Detection of pulmonary nodules by C-arm CT using a phantom lung: comparison with CT.

BACKGROUND: Evaluation of detection of lung nodules by C-arm CT (CACT) is important before this procedure can be used to guide percutaneous lung interventions. PURPOSE: To compare the efficacy of CACT with CT in the detection of pulmonary nodules using a phantom lung. MATERIAL AND METHODS: A phantom lung containing 12 phantom nodules in four sizes (5 mm/8 mm/10 mm/12 mm) and three CT values (one solid nodule, +100 HU; two ground glass nodules, -630 and -800 HU) was used. Six sessions of CACT (slice thickness 4.5 mm) and CT (slice thickness 5 mm) were performed. In each session, the locations of nodules were arbitrarily changed in the phantom. Three radiologists assessed the detection of a total of 72 nodules. Statistical analysis was performed for the sensitivity and positive predictive value of lung nodules between CACT and CT by the McNemar test and paired t-test (P < 0.05). RESULTS: Sensitivity did not differ between CACT and CT, respectively (reader 1, 82% vs. 88%, P = 0.22; reader 2, 82% vs. 78%, P = 0.37; reader 3, 79% vs. 83%, P = 0.48). For nodules of 8 mm or larger, the sensitivity increased for each reader and showed no significant difference between CACT vs. CT. The positive predictive value did not differ between CACT and CT. CONCLUSION: In this phantom study, CT and CACT show similar sensitivity for the detection of pulmonary nodules. CACT could be used in percutaneous interventional procedures in the lungs.

Adaptive statistical iterative reconstruction technique for pulmonary CT: image quality of the cadaveric lung on standard- and reduced-dose CT.

RATIONALE AND OBJECTIVES: To evaluate thin-section computed tomography (CT) images of the lung reconstructed using adaptive statistical iterative reconstruction (ASIR) on standard- and reduced-dose CT. MATERIALS AND METHODS: Eleven cadaveric lungs were scanned by multidetector-row CT with two different tube currents (standard dose, 400 mA; reduced dose, 10 mA). The degree of ASIR was classified into six different levels: 0% (non-ASIR), 20%, 40%, 60%, 80%, and 100% (maximum-ASIR). The ASIR (20%, 60%, and 100%) images were compared with the ASIR (0%) images and assessed visually by three independent observers for image quality using a 7-point scale. The evaluation items included abnormal CT findings, normal lung structures, and subjective visual noise. The median scores assigned by the three observers were analyzed statistically. Quantitative noise was calculated by measuring the standard deviation in a circular region of interest on each selected image of ASIR (0%-100%). RESULTS: On standard-dose CT, the overall image quality significantly improved with increasing degree of ASIR (P ≤ .009, Wilcoxon signed-ranks test with Bonferroni correction). As ASIR increased, however, intralobular reticular opacities and peripheral vessels tended to be obscure. On reduced-dose CT, the overall image quality of ASIR (100%) was significantly better than that of ASIR (20%) (P ≤ .009). As ASIR increased, however, intralobular reticular opacities tended to be obscure. Using ASIR significantly reduced subjective and quantitative image noise on both standard- and reduced-dose CT (P < .001, Bonferroni/Dunn's method). CONCLUSION: ASIR improves the image quality by decreasing image noise. Maximum-ASIR may be needed for improving image quality on highly reduced-dose CT. However, excessive ASIR may obscure subtle shadows.

Multidetector CT of the lung: image quality with garnet-based detectors.

PURPOSE: To evaluate the image quality of both standard- and reduced-dose computed tomography (CT) by comparing multidetector CT with garnet-based detectors with multidetector CT with conventional detectors. MATERIALS AND METHODS: The study was approved by the internal ethics review board. Informed consent was obtained. Eleven cadaveric lungs inflated and fixed by using the Heitzman method were scanned by using both CT with garnet-based detectors and CT with conventional detectors. Tube current was 400 mA for standard-dose and 10 mA for reduced-dose CT, and voltage was 120 kVp. Either normal scan mode with 984 views (conventional and garnet-based detectors) or high-resolution mode with 2496 views was used. Image quality at conventional-detector CT and garnet-based-detector CT in all modes was graded by two independent observers with a five-point scale. The evaluation items included normal lung structures, subjective visual noise, and abnormal CT findings. Quantitative image noise measurements were calculated by measuring the standard deviations in a circular region of interest on each selected image. RESULTS: At standard-dose CT, image quality at CT with garnet-based detectors (high-resolution mode) was significantly improved (P < .001, Tukey-Kramer). However, there was no significant difference between quantitative image noise measurements (P > or = .24). At reduced-dose CT, only noise differed significantly, with both subjective visual noise and quantitative image noise measurements significantly greater at CT with garnet-based detectors (high-resolution mode) (P < or = .01). There was no significant difference in image quality except for noise between conventional-detector CT and garnet-based-detector CT (P > or = .06). CONCLUSION: The image quality of standard-dose garnet-based-detector CT (high-resolution) was significantly improved. Although highly reduced-dose garnet-based-detector CT (high-resolution mode) provided more image noise, overall image quality was not different between conventional-detector CT and garnet-based-detector CT.

CT-guided percutaneous cutting needle biopsy of thymic epithelial tumors comparison to the accuracy of computed tomographic diagnosis according to the world health organization classification.

RATIONALE AND OBJECTIVES: The aims of this study were to compare diagnostic accuracy between computed tomography (CT)-guided percutaneous cutting needle biopsy (PCNB) and surgery or open biopsy for thymic epithelial tumors in accordance with the World Health Organization (WHO) classification and to evaluate computed tomographic diagnosis additionally. MATERIALS AND METHODS: Subjects were 20 patients (11 men, 9 women) in whom CT, CT-guided PCNB, and surgery had been performed for anterior mediastinal tumors. All diagnoses of both CT-guided PCNB and surgery or open biopsy were made in accordance with the WHO classification. Computed tomographic diagnoses were performed by two radiologists on the basis of radiologic characteristics previously reported according to the simplified WHO classification (types A and AB, type B1, types B2 and B3, and thymic carcinoma). The concordance of the WHO classification or the simplified WHO classification between the diagnosis on either CT or CT-guided PCNB and that on surgery was evaluated using the weighted kappa statistic. RESULTS: The histologic classifications on the basis of surgical resection specimens were as follows: type A, n = 3; type AB, n = 5; type B1, n = 3; type B2, n = 4; type B3, n = 4; and thymic carcinoma, n = 1. The overall concordance with the diagnosis according to the WHO classification established using CT-guided PCNB specimens (weighted kappa = 0.757) was higher than that using computed tomographic diagnosis (weighted kappa = 0.437). CONCLUSION: CT-guided PCNB is a technique with good concordance of the WHO classification of thymic epithelial tumors between the diagnoses of surgery or open biopsy.

A feasibility and efficacy study on bronchoscopy with a virtual navigation system.

OBJECTIVE: The purpose of this study was to evaluate the diagnostic accuracy of bronchoscopy with virtual bronchoscopy (VB) navigation. PATIENTS AND METHODS: Forty-two consecutive patients with ≤30-mm diameter peripheral pulmonary shadows or lesions that were difficult or impossible to identify on plain x-rays regardless of size were selected. Before bronchoscopy, multidetector computed tomography was performed, and VB images were created. Then, ultrathin bronchoscopy was carried out with VB navigation followed by conventional 4-mm diameter bronchoscopy. RESULTS: Thirty-seven patients were finally enrolled. There were 12 re-examination cases in which earlier conventional bronchoscopy was nondiagnostic. Diagnosis was established in 28 patients (25 malignant and 3 benign lesions), with the diagnostic yield of 75.7% (sensitivity 86.2%, specificity 62.5%, positive predictive value 89.3%, negative predictive value 55.6%, and accuracy 81.1%). The diagnostic yield for lesions of diameter ≤20 mm, 21 to 30 mm, and ≥31 mm were 76.9%, 76.5%, and 71.4%, respectively, for all patients, and 90%, 84.6%, and 83.3%, respectively, for malignant cases. The diagnostic yield in patients who underwent re-examinations was 83.3% for all cases and 88.9% for malignant cases. Three patients had completely invisible shadows on x-ray fluoroscopy and all of these cases were diagnosed on bronchoscopy. No statistically significant difference was observed between the diagnostic yield for the different sizes, pathology, and number of bronchoscopies required. CONCLUSIONS: Bronchoscopy with VB navigation yielded high accuracy even for small peripheral pulmonary nodules and shadows that were difficult or impossible to identify on plain x-rays, even for patients in whom earlier conventional bronchoscopy was nondiagnostic.