Content-based retrieval of HRCT_Depeursinge

Content-based retrieval and analysis of HRCT images from patients with interstitial lung diseases Adrien Depeursinge1, Alejandro Vargas¹, Alexandra Platon², Antoine Geissbuhler1, Pierre-Alexandre Poletti² and Henning Müller1,3 1 Medical Informatics Service, Geneva University Hospitals and University of Geneva, Geneva, Switzerland ² Service of Emergency Radiology, Geneva University Hospitals and University of Geneva, Geneva, Switzerland 3 Business Information Systems, University of Applied Sciences Sierre, Sierre Switzerland Abstract: This paper presents a computer–aided diagnosis (CAD) system to assist the radiologist in diagnosing interstitial lung diseases (ILDs) by retrieving similar cases.

Introduction The interpretation of high-resolution computed tomography (HRCT) of the chest from patients with interstitial lung diseases (ILD) is often challenging with numerous differential diagnoses [1]. We developed computer tools and created a database (133 cases representative of the 7 most frequent ILDs) to assist the radiologist in the diagnosis workup of ILDs (see Figure 1). Automatic detection and characterization of abnormal pulmonary tissue in HRCT images as well as retrieval of similar cases were implemented based on the visual information contained in the images along with a set of clinical parameters describing the clinical state of the patient.

Methods Advanced image processing algorithms and artificial intelligence techniques allowed automatic detection rates of 75.1% among fives type of lung tissue (including healthy) with a leave-one-patientout cross validation of 69 image series from patients affected with ILDs. The recognition rate is obtained with an experimental setup that is faithfully similar to actual clinical situations [2]. Similar cases are retrieved based on the volumes of abnormal tissues detected as well as subsets of clinical parameters.

Multimedia library of ILD cases: Each case of the image library contains annotated HRCT images series and 99 clinical parameters.

133 cases and 92 HRCT image series representing 7 of the most frequent ILDs have been captured.

All cases have a diagnosis confirmed by biopsy or lung washing.

Automatic 3D categorization of the lung tissue: Texture analysis is based on grey-level histograms in Hounsfield Units (HU) [3] and a custom-tailored wavelet transform [4]. A Support Vector Machine (SVM) classifier is used to predict the class of lung tissue from texture features. Classification results are displayed in three dimensions to the clinician.

Content-based retrieval of similar cases: An inter-case multimodal similarity measure based on the volumes of each class of lung tissue as well as clinical parameters were used for retrieval. Full details (including annotated image series) of the retrieved cases can be viewed with a web-based interface.

Results The automatic recognition of abnormal lung tissue provides a draft overview of the image series that constitutes a second opinion with reliability assessment. In addition, image-based retrieval of similar cases enables advanced browsing of large repositories of ILD cases [5].

3D lung tissue categorization: Table 1 shows the confusion matrix of the classification of the lung tissue patterns obtained with a leave-one-patient-out (LOPO) cross-validation of 69 cases. Recurrent confusions between the healthy and micronodules patterns are observed. Integration of clinical parameters allowed significant improvements of the classification accuracy [6].

Table 1 : Confusion matrix of the classifications of the various tissue types.

Multimodal case-based retrieval: The mean retrieval precisions P at ranks 1, 5, 10 and at rank equal to the number of instances of the diagnosis using LOPO cross-validation are listed in Table 2.

Table 2 : Performance evaluation of case-based retrieval for various interstitial lung diseases.

It can clearly be seen that the accuracy depends strongly on the diagnosis. Sarcoidosis and Fibrosis can easily be distinguished, whereas PCP has fairly low values.

Conclusions Image-based diagnosis aid tools for ILDs including a multimedia database, automatic categorization of the lung tissue and retrieval of similar cases are available for evaluation to clinicians at the Emergency Radiology Service of the HUG. The recognition rate and retrieval precisions obtained with LOPO cross-validation are similar to actual clinical situations. The automatic recognition of abnormal lung tissue provides a draft overview of the image series that constitutes a second opinion with reliability assessment. Image-based retrieval of similar cases enables advanced browsing of large repositories of ILD cases.

Future work To reduce false detections of micronodules, using SVM with asymmetric margins is on the plan. Using low-level texture features for content-based retrieval of similar cases is another possibility to enhance the performance. Our position in HUG allows keeping on identifying problems and benefits of the CAD in clinical routine at the Emergency Radiology service.

Acknowledgements This work was supported by the Swiss National Science Foundation (FNS) with grant 200020– 118638/1, the equalization fund of Geneva University Hospitals and University of Geneva (grant 05– 1 9–II) and the EU 6th Framework Program in the context of the KnowARC project (IST 032691).

References [1] T. E. King. Approach to the adult with interstitial lung disease. UpToDate, October, 2008.

[2] Dundar, M., Fung, G., Bogoni, L., Macari, M., Megibow, A., Rao, B.: A methodology for training and validating a cad system and potential pitfalls. International Congress Series 1268 (2004) 1010–1014 CARS 2004 – Computer Assisted Radiology and Surgery. Proceedings of the 18th International Congress and Exhibition.

[3] P. G. Hartley, J. R. Galvin, G. W. Hunninghake, J. A. Merchant, S. J. Yagla, S. B. Speakman, and D. A. Schwartz. High{resolution CT{derived measures of lung density are valid indexes of interstitial lung disease.

Journal of Applied Physiology, 76(1):271–277, January 1994.

[4] Depeursinge, A., Van De Ville, D., Unser, M., Müller, H.: Lung tissue analysis using isotropic polyharmonic Bspline wavelets. In: MICCAI 2008 Workshop on Pulmonary Image Analysis, New York, USA (September 2008) 125–134.

[5] H. Müller, N. Michoux, D. Bandon, and A. Geissbuhler. A review of content–based image retrieval systems in medicine – clinical benefits and future directions. Int. Journal of Medical Informatics, 73:1–23, 2004.

[6] Depeursinge, A., Racoceanu, D., Iavindrasana, J., Cohen, G., Platon, A., Poletti, P.A., Müller, H.: Fusing visual and clinical information for lung tissue classification in HRCT data. Journal of Artificial Intelligence in Medicine (2009 – to appear).

Correspondence Adrien Depeursinge Geneva University Hospitals and University of Geneva, Service of Medical Informatics, Rue Gabrielle-Perret-Gentil 4, CH-1211 Geneva 14, Switzerland adrien.depeursinge@sim.hcuge.ch, http://www.sim.hcuge.ch/medgift/ Tel ++ 41 22 372 8875 Fax ++ 41 22 372 8879

1

http://www.knowarc.eu