PLoS One. 2015 Sep 30; 10(9):e0139511. doi: 10.1371/journal.pone.0139511.

Detecting Lung Diseases from Exhaled Aerosols: Non-Invasive Lung Diagnosis Using Fractal Analysis and SVM Classification.

Jinxiang Xi1, Weizhong Zhao2,3, Jiayao Eddie Yuan1, JongWon Kim4, Xiuhua Si5, Xiaowei Xu6

1School of Engineering and Technology, Central Michigan University, Mount Pleasant, Michigan, United States of America.

2College of Information Engineering, Xiangtan University, Xiangtan, Hunan Province, China.

3Division of Bioinformatics and Biostatistics, National Center for Toxicological Research, Jefferson, Arkansas, United States of America.

4College of Engineering, University of Georgia, Athens, Georgia,  United States of America.

5Department of Mechanical Engineering, California Baptist University, Riverside, California, United States of America.

6Department of Information Science, University of Arkansas, Little Rock, Arkansas, United States of America.



Lung cancers’ symptomless development makes their early detection extremely difficult.  In this study, we introduced a novel lung diagnostic technique that used exhaled aerosols to determine the severity and location of lung diseases. Our hypothesis was that exhaled aerosols disclose significant information pertaining to lung structure; their distribution is like a fingerprint specific to the lungs which could be analyzed by a computer algorithm to identify what the disease is, where it is located, and how wide-spread it is.  The exhaled aerosol fingerprints were first quantified using fractal analysis and subsequently correlated with lung disease using data-mining algorithms.  A computerized database has shown the method can diagnose obstructive lung disease with an accuracy of 99% in four specific cases.  This study has, for the first time, statistically linked exhaled aerosol patterns with their underlying diseases.

The proposed aerosol-based lung diagnostic method offers several advantages over current methods.  First, compared to current diagnostic tools such as chest radiography, CT, and biopsy, the proposed aerosol breath test is non-invasive, cost-efficient, easy to perform, and providing real-time feedback.  Second, compared to other breath tests in development, such as the electronic nose, exhaled breath condensate, and aerosol bolus dispersion, this method has the unique potential to locate a disease, which is highly desirable for personalized site-specific drug delivery. Third, in contrast to other tumor-classification studies that require the images of cancerous tissue samples, the new algorithm in this study only requires an exhaled aerosol pattern, which can be obtained easily.

PMID: 26422016



We developed a Fractal-Data-Mining algorithm that can detect a lung structural abnormality, determine the severity of the disease, and monitor the disease growth or therapeutic outcome of the treatment [1,2,3].  This algorithm consists of two steps: image feature extraction using sub-regional fractal analysis and data classification using a support vector machine (SVM). A schematic diagram is shown in Fig. 1A.

Lung disease models: The performance of this algorithm was tested in an image-based mouth-throat airway model developed by Xi and Longest [4].  This model geometry was modified to produce the three asthmatic models by progressively decreasing the diameters of two segmental bronchioles, as illustrated in Fig. 1B.  The information on location, size, and airway blockage rate of the three asthmatic models is listed in Fig. 1B.



jx fig1

Figure 1. Aerosol-based lung diagnostic method: (A) flow chart, (B) asthmatic models with three levels of constrictions, and (C) image classification using a support vector machine (SVM) algorithm. A classification accuracy of 99% for four cases of asthma was achieved based on a database containing 324 digital aerosol images.


Image database development: Computational fluid-particle dynamics (CFPD) modeling was employed to simulate aerosol inhalation and exhalation by a patient.  The resulting exhaled aerosol images were visualized in terms of both particle locations and concentration distributions (Fig. 1C, upper panel).  A database containing 324 images was developed by considering four asthma cases under the influence of varying sources of uncertainties, such as particle size, inhalation flow rate, and geometry variation in the oral cavity and the throat.

Image quantification: It may appear straightforward that different lung structures generate different exhaled aerosol patterns.  However, to differentiate them in a quantitative manner presents many challenges.  First, the exhaled aerosol patterns are very complex.  Second, the discrepancy between different aerosol patterns can be too subtle to be discerned visually.  To characterize such patterns, the images were divided into a 6×6 lattice and fractal dimension analysis was performed on each grid, leading to a 36-variable vector for each image (Fig 1C, lower panel).

Classification: In principle, one disease contains certain unique features in the exhaled aerosols despite uncertainties such as aerosol size, breathing rates, and body positions.  Data mining techniques were used to statistically identify such features as indicators for disease detection and localization.  A multi-class SVM algorithm and 10-fold cross-validation were used for the classification analysis. The database was randomly divided into 10 subsets and the SVM algorithm was run 10 times. In each run, 9 subsets were used as a training set to obtain an SVM classifier, and one subset was used for testing.  After 10 runs, each subset was tested once and trained 9 times.  Following this method, a total prediction accuracy of 99% was achieved, with only three misdiagnosed samples from the 324 sample database (Fig. 1D).



  1. Xi J, Kim J, Si XA, Zhou Y (2013) Diagnosing obstructive respiratory diseases using exhaled aerosol fingerprints: A feasibility study. J Aerosol Sci 64: 24-36.
  2. Xi J, Si XA, Kim J, Mckee E, Lin E-B (2014) Exhaled aerosol pattern discloses lung structural abnormality: a sensitivity study using computational modeling and fractal analysis. PLoS ONE 9: e104682.
  3. Xi J, Kim J, Si XA, Mckee E, Corley RA, Kabilan S, et al. (2015) CFD modeling and image analysis of exhaled aerosols due to a growing bronchial tumor: towards non-invasive diagnosis and treatment of obstructive respiratory diseases. Theranostics 5: 443-455.
  4. Xi J, Longest PW (2007) Transport and deposition of micro-aerosols in realistic and simplified models of the oral airway. Ann Biomed Eng 35: 560-581.



Multiselect Ultimate Query Plugin by InoPlugs Web Design Vienna | Webdesign Wien and Juwelier SchönmannMultiselect Ultimate Query Plugin by InoPlugs Web Design Vienna | Webdesign Wien and Juwelier Schönmann