Pulmonary Textures Classification via a Multi-Scale Attention Network.

Precise classification of pulmonary textures is crucial to develop a computer aided diagnosis (CAD) system of diffuse lung diseases (DLDs). Although deep learning techniques have been applied to this task, the classification performance is not satisfied for clinical requirements, since commonly-used deep networks built by stacking convolutional blocks are not able to learn discriminative feature representation to distinguish complex pulmonary textures. For addressing this problem, we design a multi-scale attention network (MSAN) architecture comprised by several stacked residual attention modules followed by a multi-scale fusion module. Our deep network can not only exploit powerful information on different scales but also automatically select optimal features for more discriminative feature representation. Besides, we develop visualization techniques to make the proposed deep model transparent for humans. The proposed method is evaluated by using a large dataset. Experimental results show that our method has achieved the average classification accuracy of 94.78% and the average f-value of 0.9475 in the classification of 7 categories of pulmonary textures. Besides, visualization results intuitively explain the working behavior of the deep network. The proposed method has achieved the state-of-the-art performance to classify pulmonary textures on high resolution CT images.

Ultra-low-dose CT of the lung: effect of iterative reconstruction techniques on image quality.

RATIONALE AND OBJECTIVES: To compare quality of ultra-low-dose thin-section computed tomography (CT) images of the lung reconstructed using model-based iterative reconstruction (MBIR) and adaptive statistical iterative reconstruction (ASIR) to filtered back projection (FBP) and to determine the minimum tube current-time product on MBIR images by comparing to standard-dose FBP images. MATERIALS AND METHODS: Ten cadaveric lungs were scanned using 120 kVp and four different tube current-time products (8, 16, 32, and 80 mAs). Thin-section images were reconstructed using MBIR, three ASIR blends (30%, 60%, and 90%), and FBP. Using the 8-mAs data, side-to-side comparison of the four iterative reconstruction image sets to FBP was performed by two independent observers who evaluated normal and abnormal findings, subjective image noise, streak artifact, and overall image quality. Image noise was also measured quantitatively. Subsequently, 8-, 16-, and 32-mAs MBIR images were compared to standard-dose FBP images. Comparisons of image sets were analyzed using the Wilcoxon signed rank test with Bonferroni correction. RESULTS: At 8 mAs, MBIR images were significantly better (P < .005) than other reconstruction techniques except in evaluation of interlobular septal thickening. Each set of low-dose MBIR images had significantly lower (P < .001) subjective and objective noise and streak artifacts than standard-dose FBP images. Conspicuity and visibility of normal and abnormal findings were not significantly different between 16-mAs MBIR and 80-mAs FBP images except in identification of intralobular reticular opacities. CONCLUSIONS: MBIR imaging shows higher overall quality with lower noise and streak artifacts than ASIR or FBP imaging, resulting in nearly 80% dose reduction without any degradations of overall image quality.

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.

Long-term follow-up high-resolution CT findings in non-specific interstitial pneumonia.

BACKGROUND: The aims of this study were to retrospectively assess the change in findings on follow-up CT scans of patients with non-specific interstitial pneumonia (NSIP; median, 72 months; range, 3-216 months) and to clarify the correlation between the baseline CT findings and mortality. METHODS: The study included 50 patients with a histologic diagnosis of NSIP. Two observers evaluated the high-resolution CT (HRCT) findings independently and classified each case into one of the following three categories: (1) compatible with NSIP, (2) compatible with UIP or (3) suggestive of alternative diagnosis. The correlation between the HRCT findings and mortality was evaluated using the Kaplan-Meier method and the log-rank test, as well as Cox proportional hazards regression models. RESULTS: Ground-glass opacity and consolidation decreased, whereas coarseness of fibrosis and traction bronchiectasis increased on the follow-up HRCT scans, however, in 78% of cases the overall extent of parenchymal abnormalities had no change or decreased. Patients with HRCT diagnosed compatible with NSIP had a longer survival than those with HRCT findings more compatible UIP or an alternative diagnosis. On multivariate analysis, the coarseness of fibrosis alone was associated with prognosis (HR: 1.480; 95% CIs 1.100 to 1.990). CONCLUSIONS: The HRCT patterns seen in patients with a histopathologic diagnosis of NSIP progress in a variable manner. Overall disease extent may decrease over time in some, while fibrosis may progress in others. The initial HRCT diagnosis may impact survival in this group of patients.