AR: Pulmonary infection CAD shows promise for H1N1
Chest radiograph of patient with H1N1 virus. Image Source: American Journal of Roetgenology
Texture analysis may offer a quantitative method to differentiate abnormal and normal lung regions in patients with severe H1N1 infection and provide standardized, automated detection of abnormal areas, according to a study published in this month's Academic Radiology.

The clinical utility of infectious disease imaging is hampered by low specificity, limited quantification of disease burden and observer-dependent visual scoring methods, according to Jianhua Yao, PhD, from the Center for Infectious Disease Imaging and department of radiology and imaging sciences at the National Institutes of Health in Bethesda, Md., and colleagues.

These constraints became apparent during the H1N1 outbreak. “It was clinically and radiologically challenging to prognosticate and identify [severe H1N1 cases] for earlier treatment,” wrote Yao.

Yao and colleagues developed and tested a method for detecting and quantifying H1N1 pulmonary infection using computer-assisted texture analysis and support vector machines. “Texture analysis quantifies an image by identifying statistical relationships among the pixels’ densities, which can be used to identify lesions and quantify the volume of an organ manifesting those patterns associated with lesions,” the researchers explained.

Researchers studied CT datasets from 10 patients with proven H1N1 infection, 10 normal controls and 20 patients with fibrosis. The study focused on the quantitative differentiation of lesions in patients with H1N1 from visually normal lung parenchyma and comparison of texture features of H1N1 CT image datasets to patterns in the patients with normal lungs, fibrosis and other infections.

The pilot computer-aided detection (CAD) system segments the lungs, analyzes texture patterns and distinguishes normal and abnormal regions. Yao and colleagues calculated six computed texture values: density mean, density deviation, correlation, average sum, gray-level non-conformity and gray-level run emphasis.

The analysis showed consistent quantitative differentiation between abnormal regions on H1N1 CT studies versus fibrosis and consistent differentiation between normal CT data and normal-appearing regions of CT images from patients with H1N1, reported Yao and colleagues.

The program also differentiated acute infection from fibrosis, which could help evaluate infection disease onset and progression. “However, this texture analysis did not differentiate bacterial pneumonia from H1N1, which supports the known association between influenza and bacterial co-infection,” wrote Yao.

The ability of texture analysis to differentiate abnormal and normal areas of the lung parenchyma in patients with H1N1 infection could be applied in both the research and clinical settings to automate detection of abnormal regions and quantify the abnormality in proportion to lung volumes, suggested the researchers.

While acknowledging the positive results and potential utility of the technique, the authors noted, that future investigations are needed "to assess the potential of this technique as a clinical tool in diagnosing disease, quantifying severity, monitoring response to therapy and predicting outcomes.”  Yao and colleagues suggested multiple avenues for future research: an expanded case set, inclusion of more diverse etiologies of fibrosis, correlation of lesions with texture features and assessment of co-infection.