Ce diaporama a bien été signalé.
Le téléchargement de votre SlideShare est en cours. ×

Retrieval of 4D Dual Energy CT for Pulmonary Embolism Diagnosis

Publicité
Publicité
Publicité
Publicité
Publicité
Publicité
Publicité
Publicité
Publicité
Publicité
Publicité
Publicité

Consultez-les par la suite

1 sur 19 Publicité

Retrieval of 4D Dual Energy CT for Pulmonary Embolism Diagnosis

Télécharger pour lire hors ligne

Pulmonary embolism is a common condition with high short–term morbidity. Pulmonary embolism can be treated successfully but diagnosis remains difficult due to the large variability of symptoms, which are often non–specific including breath shortness, chest pain and cough. Dual energy CT produces 4–dimensional data by acquiring variation of
attenuation with respect to spatial coordinates and also with respect to the energy level. This additional information opens the possibility of discriminating tissue with specific material content, such as bone and adjacent contrast. Despite having already been available for clinical use for a while, there are few applications where Dual energy CT is currently showing a clear clinical advantage. In this article we propose to use the additional energy–level data in a 4D dataset to quantify texture changes in lung parenchyma as a way of finding parenchyma perfusion deficits characteristic of pulmonary embolism.

Pulmonary embolism is a common condition with high short–term morbidity. Pulmonary embolism can be treated successfully but diagnosis remains difficult due to the large variability of symptoms, which are often non–specific including breath shortness, chest pain and cough. Dual energy CT produces 4–dimensional data by acquiring variation of
attenuation with respect to spatial coordinates and also with respect to the energy level. This additional information opens the possibility of discriminating tissue with specific material content, such as bone and adjacent contrast. Despite having already been available for clinical use for a while, there are few applications where Dual energy CT is currently showing a clear clinical advantage. In this article we propose to use the additional energy–level data in a 4D dataset to quantify texture changes in lung parenchyma as a way of finding parenchyma perfusion deficits characteristic of pulmonary embolism.

Publicité
Publicité

Plus De Contenu Connexe

Plus par Institute of Information Systems (HES-SO) (20)

Publicité

Plus récents (20)

Retrieval of 4D Dual Energy CT for Pulmonary Embolism Diagnosis

  1. 1. Retrieval of 4D Dual Energy CT for Pulmonary Embolism Diagnosis A. Foncubierta-Rodríguez, A. Vargas, A. Platon, P.A. Poletti, H. Müller, A. Depeursinge
  2. 2. Overview • Motivation: pulmonary embolism and Dual-Energy CT Imaging • Methods: 4D texture analysis and retrieval • Results • Conclusions and future work
  3. 3. Introduction to Pulmonary Embolism • Acute Pulmonary Embolism has high mortality rates • 30% deaths 3 years after hospital discharge • 75% of deaths occur during initial hospital admission • Avoidable cause of death if treated immediately • Delays in diagnosis increase risk of death • CT appearance of the embolized lung parenchyma • Wedge-shaped regions • Simple 3D Texture Can Dual Energy CT add extra information?
  4. 4. Features of Dual Energy CT • Dual Energy CT contains 4D data • 3 spatial coordinates: x, y, z and 1 energy coordinates: e • At varying energy levels, attenuation curves behave differently for different materials Material Attenuation Coefficient vs keV 10 0 80 keV Iodine 140 keV (cm2/mg) 1 0 m(E) 1 Water 0. 1 40 50 60 70 80 90 100 110 120 130 140 Photon Energy (keV) • E.g., allows for isolating iodine components
  5. 5. Motivation • Dual Energy CT is currently underused • Lack of a clear clinical application • Limitations of displaying dimensions larger than 3 for human inspection • Computerized 4D texture analysis • Comprehensive data analysis • Clinical benefits and opportunities • Quantitative comparisons of the parenchyma between healthy and embolism • Visualization of ischemia • Content-based retrieval of cases with similar Dual Energy CT patterns as diagnosis aid
  6. 6. Methods: Dataset and Ground Truth • Dataset: • 13 patients • Image resolution • 0.83mm/voxel (axial plane) • 1mm inter-slice distance • 1.25mm slice thickness • 11 energy levels • Ground truth • Lobes manually segmented • Qanadli index per lobe
  7. 7. Methods: Qanadli index • Severity of pulmonary embolism • Computation depends on artery occlusion level • +0: no occlusion • +1: partial • +2 : complete occlusion
  8. 8. Methods: 4D Texture analysis • One 3D volume is obtained per energy level • 3D low level features are computed for each volume: • 3D difference of Gaussians wavelet coefficients • 1 to 5 scales in dyadic progression • Energy of coefficients in a 6x6x6 neighborhood (overlapping) • High level features are computed in a Leave One Patient Out cross-validation
  9. 9. Methods: 4D Texture analysis • High-level features: Bags of visual words • Leaving one patient out, the low level features are clustered into a varying number of visual words • The remaining patient is then described by the histogram of visual words for each lobe • Euclidean distance for retrieval
  10. 10. Experimental setup PARAMETERS EVALUATION • Data Sources • Retrieval of similar lobes • All 11 energy levels • Healthy: Qanadli = 0 (21) • Only 70 KeV energy level • Embolism: Qanadli > 0 (44) • Visual Words • Precision measured • 50, 100, 150 • P@1 • Wavelet scales • P@5 • 1, 2, 3, 5 scales • P@10
  11. 11. Results: 3D Data (70 KeV) Scales Visual Words P@1(%) P@5(%) P@10(%) 50 60 56 56 1 100 57 57 55 150 58 56 54 50 55 57 55 2 100 58 57 56 150 63 60 54 50 51 55 53 3 100 49 53 55 150 55 55 55 50 45 50 52 5 100 51 49 52 150 51 52 53 Mean 54.42 54.75 54.17 Standard Deviation 4.92 3.06 1.34
  12. 12. Results: 4D Data (all energy levels) Scales Visual Words P@1(%) P@5(%) P@10(%) 50 55 56 56 1 100 58 55 57 150 58 56 56 50 62 58 55 2 100 62 62 60 150 63 62 60 50 58 54 55 3 100 60 59 58 150 57 62 58 50 45 52 51 5 100 57 52 51 150 58 52 52 Mean 57.75 56.67 55.75 Standard Deviation 4.47 3.75 3.00
  13. 13. Results: Comparison 3D (70KeV) and 4D Scales Visual Words P@1(%) P@5(%) P@10(%) 50 -5 0 0 1 100 1 -2 2 150 0 0 2 50 7 1 0 2 100 4 5 4 150 0 2 6 50 7 -1 2 3 100 11 6 3 150 2 7 3 50 0 2 -1 5 100 6 3 -1 150 7 0 -1 22 out of 36 Experiments where 4D performs better than 3D 6 out of 36 Experiments where 3D performs better than 4D
  14. 14. Best configurations (P@1) 4D Data. 2 scales, 150 VW 3D Data. 2 scales, 150 VW Healthy Embolism Healthy Embolism Healthy 47.6% 52.4% Healthy 52.4% 47.6% Embolism 29.5% 70.5% Embolism 31.8% 68.2% Same configuration: 2 scales, 150 VW ~63% Different data source
  15. 15. Best configurations (P@5) 4D Data. 2 scales, 100 VW 4D Data. 2 scales, 150 VW Healthy Embolism Healthy Embolism Healthy 43.8% 56.2% Healthy 43.8% 56.2% Embolism 29.5% 70.5% Embolism 29.9% 70.1% 4D Data. 3 scales, 150 VW Only 4D Data Healthy Embolism Healthy 44.8% 55.2% Very small Embolism 29.1% 70.9% ~62% differences Embolism cases are better retrieved
  16. 16. Best configurations (P@10) 4D Data. 2 scales, 100 VW 4D Data. 2 scales, 150 VW Healthy Embolism Healthy Embolism Healthy 42.4% 57.6% Healthy 40% 60% Embolism 31.1% 68.9% Embolism 31.8% 68.9% Only 4D Data ~60% Embolism cases are better retrieved
  17. 17. Conclusions & Future Work • 4D texture analysis of Dual Energy is relevant for the characterization pulmonary embolism • Best performing configurations suggest • 4D data • 2 wavelet scales • 150 visual words, but more can lead to better results • Future work • Complete the dataset with more control patients • New dataset with 19 patients with embolism and 8 control already available
  18. 18. Conclusions & Future Work • Investigate other low-level features • Grey-level histograms in Hounsfield Units • Non-isotropic wavelets • 2 first principal 5 4 components: 3 2 x healthy 1 PC2 o embolism 0 -1 -2 -3 -4 -12 -10 -8 -6 -4 -2 0 2 4 6 PC1
  19. 19. Thank you for your attention! A. Foncubierta-Rodríguez, A. Vargas, A. Platon, P.A. Poletti, H. Müller and A. Depeursinge. “Retrieval of 4D Dual Energy CT for Pulmonary Embolism Diagnosis” in: Medical Content-based Retrieval for Clinical Decision Support, Nice, France, 2012 Collaborations: • Prof. Antoine Geissbuhler, University Hospitals of Geneva • Dr. Pierre-Alexandre Poletti, University Hospitals of Geneva • Prof. Dimitri Van de Ville, EPFL

×