Research ArticleComputational Biology

Automated identification of abnormal respiratory ciliary motion in nasal biopsies

See allHide authors and affiliations

Science Translational Medicine  05 Aug 2015:
Vol. 7, Issue 299, pp. 299ra124
DOI: 10.1126/scitranslmed.aaa1233
  • Fig. 1. Properties of CM.

    (A) Schematic (hand-drawn) diagrams of CM subtypes to aid clinical diagnosis. (B) Stacked frames indicate still frames of the video of the CM biopsy, and the black box indicates the ROI selected by the clinician. (C) Yellow arrows on the images from (B) indicate direction and magnitude of optical flow for a small region of the video for each pair of frames. (D) Changes in the optical flow are used to compute the elemental components. Red arrows, optical flow at frame t; green arrows, optical flow at frame t + 1; blue arrows, optical flow at frame t + 2. (E) Elemental components of rotation (top left), deformation (top right, bottom right), and divergence (bottom left; excluded from analysis), shown here in a template form. Deformation is a vectorial quantity requiring two templates for measurement.

  • Fig. 2. Digital representations of pixels in ciliary biopsy videos.

    (A) Single frame of a video of normal and abnormal CM with three pixels identified: blue (proximal to cell wall), red (distal from cell wall), and black (background). See movies S1 to S9 for examples of normal and abnormal CM. (B) Time series of gray-level pixel intensities over 100 frames at each of the three respective pixel locations in (A). (C) Time series of rotation over 100 frames at each of the respective pixel locations in (A). (D) Time series of deformation amplitude over 100 frames at each of the respective pixel locations in (A). a.u., arbitrary units.

  • Fig. 3. CM AR representations.

    (A) Top five principal components of CHP rotation data and the percentage of the overall variance in the CM data explained by each component. The top q principal components are used to compute the AR motion parameters. (B) One-dimensional rotation signal from a single pixel of normal (left) and abnormal (right) CM as the original signal (darkest blue/red) is reconstructed using an increasing number of principal components. Darker lines indicate larger q (shown: q = 2, q = 5, and q = 10).

  • Fig. 4. CM histogram representations and results.

    (A) Time domain histograms of ciliary rotation and deformation magnitudes from normal (blue) and abnormal (red) CM. The time series in Fig. 2 was projected onto the vertical axis and normalized. (B) Frequency domain histograms of ciliary rotation and deformation time series from normal and abnormal CM. A fast Fourier transform was used on the rotation and deformation time series, the dominant frequency at each pixel was computed from the Fourier response, and histograms of these frequencies for rotation and deformation were plotted.

  • Fig. 5. CM AR model results.

    (A) CM is visualized using the first three dimensions of the subspace of the AR model for normal and abnormal CM. This motion is governed by the AR coefficients. Passage of time is indicated by hue, darkening with each discrete time increment. (B) Histograms show the distributions of values taken by normal and abnormal AR motion in each of the first three PCA dimensions.

  • Table 1. Description and breakdown of data sets.

    Both data cohorts consisted of a mix of patients for whom their CM was assessed manually as either normal or abnormal; these assessments were used as ground truth for validating our framework. Multiple video samples were generated for each patient, and from these videos, multiple ROIs were selected for analysis. CHD, congenital heart disease.

    DiagnosisIndividualsVideosROIs
    Children’s Hospital of Pittsburgh
      Healthy controls2776114
      PCD controls53896
      CHD/TGA with abnormal CM1756121
      Total49170331
    Children’s National Medical Center
      PCD controls42558
      Heterotaxy with normal CM1765139
      Heterotaxy with abnormal CM103165
      Total31121262
  • Table 2. Classification accuracy, sensitivity, and specificity on both data cohorts, with each proposed method.

    We compare the performance of our methods to two baseline methods: classification using the histogram method on the gray-level pixel intensities in lieu of computing optical flow, and using CBF.

    MethodData
    set
    Accuracy
    (%)
    Sensitivity
    (%)
    Specificity
    (%)
    Proposed methods
      HistogramCHP93.8895.2492.86
      AR (rotation)CHP88.6480.0095.83
      AR (deformation)CHP86.3676.1995.65
      HistogramCNMC86.6791.6783.33
      AR (rotation)CNMC83.3383.3383.33
      AR (deformation)CNMC70.0059.09100.00
    Baseline methods
      Histogram
    (intensities)
    CHP72.7363.1680.00
      CBFCHP52.2735.0058.33

Supplementary Materials

  • www.sciencetranslationalmedicine.org/cgi/content/full/7/299/299ra124/DC1

    Materials and Methods

    Fig. S1. Aggregate optical flow displacement in CHP data cohort.

    Fig. S2. Pixel selection in an ROI.

    Fig. S3. Breakdown of digital nasal biopsy video data sets.

    Fig. S4. CM classification pipeline.

    Fig. S5. Classification confidence as a function of ROIs per patient.

    Fig. S6. Web site proof-of-concept screenshots.

    Fig. S7. Pairwise angles between principal components of CM in AR models.

    Fig. S8. CM classification results of parameter scanning.

    Table S1. Constant parameters used throughout this study.

    Movie S1. Example of normal CM of nasal biopsy from control.

    Movie S2. Example of abnormal CM of nasal biopsy from PCD patient.

    Movie S3. Example of abnormal asynchronous and wavy CM.

    Movie S4. Example of abnormal CM with incomplete stroke.

    Movie S5. Example of abnormal CM with asynchronous beat and incomplete stroke.

    Movie S6. Example of a video capture artifact of extraneous tissue motion.

    Movie S7. Example of a video capture artifact of poor camera focus.

    Movie S8. Example of a false-negative prediction.

    Movie S9. Example of a false-positive prediction.

    Reference (42)

  • Supplementary Material for:

    Automated identification of abnormal respiratory ciliary motion in nasal biopsies

    Shannon P. Quinn, Maliha J. Zahid, John R. Durkin, Richard J. Francis, Cecilia W. Lo,* S. Chakra Chennubhotla*

    *Corresponding author. E-mail: cel36{at}pitt.edu (C.W.L.); chakracs{at}pitt.edu (S.C.C.)

    Published 5 August 2015, Sci. Transl. Med. 7, 299ra124 (2015)
    DOI: 10.1126/scitranslmed.aaa1233

    This PDF file includes:

    • Materials and Methods
    • Fig. S1. Aggregate optical flow displacement in CHP data cohort.
    • Fig. S2. Pixel selection in an ROI.
    • Fig. S3. Breakdown of digital nasal biopsy video data sets.
    • Fig. S4. CM classification pipeline.
    • Fig. S5. Classification confidence as a function of ROIs per patient.
    • Fig. S6. Web site proof-of-concept screenshots.
    • Fig. S7. Pairwise angles between principal components of CM in AR models.
    • Fig. S8. CM classification results of parameter scanning.
    • Table S1. Constant parameters used throughout this study.
    • Legends for movies S1 to S9
    • Reference (42)

    [Download PDF]

    Other Supplementary Material for this manuscript includes the following:

    • Movie S1 (.mov format). Example of normal CM of nasal biopsy from control.
    • Movie S2 (.mov format). Example of abnormal CM of nasal biopsy from PCD patient.
    • Movie S3 (.mov format). Example of abnormal asynchronous and wavy CM.
    • Movie S4 (.mov format). Example of abnormal CM with incomplete stroke.
    • Movie S5 (.mov format). Example of abnormal CM with asynchronous beat and incomplete stroke.
    • Movie S6 (.mov format). Example of a video capture artifact of extraneous tissue motion.
    • Movie S7 (.mov format). Example of a video capture artifact of poor camera focus.
    • Movie S8 (.mov format). Example of a false-negative prediction.
    • Movie S9 (.mov format). Example of a false-positive prediction.

    [Download Movies S1 to S9]

Navigate This Article