Detection of pulmonary nodules in 3D CT data

Martin Dolejší1, Jan Kybic1, Michal Polovinčák2, Stanislav Tůma2

1 CMP, Department of Cybernetics, Czech Technical University in Prague, Prague, Czech Republic
2 Department of Imaging Methods, Faculty Hospital in Motol, Prague, Czech Republic

Motivation

We are developing system for automated detection of lung nodules from CT scans.

Examples of system output. Red-true nodules, white-false positives, blue-possible nodule missed by human.

3D visualization of input data. You can also see the video.

  • Input: 3D CT image of lung region.
  • Task: locate all pulmonary nodules.
  • Different shape of nodules:
    • Juxta-pleural nodules adjected to pleura with worm shape.
    • Parenchymal nodules located in lung parenchyma with spherical shape.
Example of juxta-pleural nodule.
Example of parenchymal nodule.
  • Lung cancer is one of the leading causes of cancer related death (31% of males and 26% of females) in the USA and other developed countries.
  • Most cases of lung cancer deaths are due to smoking (www.cancer.org).
  • Early detection of lung cancer helps to improve the five-year survival rate significantly.

Image from Cancer Facts and Figures 2008

Proposed methods

Step I - Nodule candidates finding

Juxta-pleural nodules

Parenchymal nodules
  • Methods of mathematical morphology are used.
  • Only objects on the lung borders detected.
  • Only object smaller than29 mm2 detected.
  • Scale sensitive filtration is used.
  • Quasi spherical objects detected.

Step II - Nodule candidates classifying

  • Classification have to decrease number of false positives result.
  • Have not to lower the sensitivity.
  • Asymmetric AdaBoost classifier:
    • It can handle features of different values.
    • Performs feature selection.
    • Is cost sensitive.
  • Support Vector Machines (SVM) classifier:
    • Kernel based method.
    • Maximize margin.
    • Allows cost sensitive learning.

Experiments

Candidates detection

  • Different parameters of detectors was chosen to construct a ROC curve of detecting algorithm.
  • The working point was selected to 95.5% of sensitivity with 16,9 FPs/slice.

Candidates classification

Asymmetric AdaBoost

Support Vector machines (SVM)
  • Quick learning.
  • Quick classifying.
  • One parameter to set.
  • Slow learning.
  • Slow classifying.
  • A lot of changeable parameters.

Conclusion

  • Novel algorithms for parenchymal nodules detection were implemented and tested.
  • From the experiments, we conclude that:
    • The sensitivity of the system is comparable or better than in other state of the art methods.
    • The number of false positive results has to be lowered.
    • The Asymmetric AdaBoost performs is more accurate and faster than SVM.
    • The algorithm is dataset sensitive.
  • We published the annotated dataset Lung TIME to allow objective comparison of different systems.

Publications

  • M. Dolejsi, J. Kybic, M. Polovincak, and S. Tuma, "Reducing false positive responses in lung nodule detector system by Asymmetric AdaBoost," in Proceedings of 2008 IEEE International Symposium on Biomedical Imaging: From Nano to Macro. IEEE, May 2008, pp. 656-659. [ BIB ] [ PDF] ]

  • S. Tuma, J. Neuwirth, M. Dolejsi, J. Kybic, K. Danickova, M. Polovincak, J. Sanda, E. Cumlivska, and J. Malis, "Moznost snizovani davky ionizujiciho zareni pri vysetreni nadorovych onemocneni plic u deti a dorostu." in Cesko-slovenska Pediatrie, vol. 62, no. 5. Praha, Czech Republic: Ceska lekarska spolecnost J. E. Purkyne, June 2007, pp. 357-358.

  • M. Dolejsi and J. Kybic, "Automatic two-step detection of pulmonary nodules," in Proceedings of SPIE, ser. Medical Imaging 2007: Computer-Aided Diagnosis, M. L. Giger and N. Karssemeijer, Eds., vol. 6514. SPIE, February 2007, pp. 3j-1-3j-12. [ BIB ] [ PDF] ]

  • M. Dolejsi, "Detection of pulmonary nodules from CT scans," MSc Thesis CTU-CMP-2007-05, Center for Machine Perception, K13133 FEE Czech Technical University, Prague, Czech Republic, January 2007. [ BIB ] [ PDF] ]

  • J. Neuwirth, M. Polovincak, J. Kybic, S. Tuma, E. Cumlivska, V. Suchanek, T. Adla, V. Hlavac, and M. Dolejsi, "Solitarni a mnohocetne plicni uzly - analyza morfologickych vlastnosti na prukaz jejich etiologie, pravidla sledovani malych uzlu, jejich detekce pomoci semiautomaticke analyzy CT obrazu - vlastni zkusenosti a prehled literatury," Ceska radiologie, vol. 60, no. 5, pp. 311-320, March 2006.

  • M. Dolejsi and J. Kybic, "Detection of pulmonary nodules in CT scans," in Analysis of Biomedical Signals and Images - Proceedings of Biosignal 2006,J. Jan, J. Kozumplik, and I. Provaznik, Eds. Brno, Czech Republic: VUTIUM Press, June 2006, pp. 251-253. [ BIB ]