[home] [Personal Program] [Help]

---

Session: Imaging - Cardiac System
Session starts: Thursday 24 January, 13:30
Presentation starts: 14:45
Room: Lecture room 559

---

Anne Saris ()
Maartje Nillesen ()
Richard Lopata ()
Chris de Korte ()

Abstract:
Automated segmentation of three-dimensional (3D) echocardiographic images in patients with congenital heart disease is challenging, because of poor contrast between blood and cardiac tissue locally. Incorporation of a priori knowledge of anatomy is undesirable for patients with congenital malformations. Cardiologists mentally incorporate movement of the heart, using temporal coherence of structures to resolve ambiguities. Therefore, we investigated the merit of temporal cross-correlation for automated segmentation over the entire cardiac cycle. A 3D iterative cross-correlation algorithm [1] was extensively examined with respect to signal choice (envelope or radio-frequency (RF) data) and window size, in order to obtain optimal contrast of maximum cross-correlation (MCC) values between blood and cardiac tissue in all phases of the cardiac cycle. Both contrast (CNR and Overlap) and boundary-gradient (Acutance) were quantified. Resulting optimal MCC-values were used as additional external force in a deformable model approach [2] to segment the left ventricular cavity in 3D echocardiographic images, in entire systolic phase. MCC-values were tested against and combined with, adaptive filtered, demodulated RF-data. Segmentation results were compared with manually segmented volumes using a 3D Dice similarity index (3D SI). Results in 3D pediatric echocardiographic images sequences (n = 5) demonstrate that the use of envelope data outperformed RF-data in terms of optimal blood-myocardium Acutance, Overlap and CNR. The use of a relatively small axial window (0.7 – 1.25 mm) resulted in optimal contrast and boundary gradient between the two tissues. Incorporation of MCC-values, either alone or in combination with adaptive filtered, demodulated RF-data, improved automated segmentation of the left ventricular cavity (n = 4). When MCC-values were used as external force in the deformable model, the 3D SI increased in 75% of the cases (average SI increase: 0.71 to 0.82). Results might be further improved by optimizing MCC-contrast locally, in regions with low blood-tissue contrast. Reducing underestimation of the endocardial volume due to MCC processing scheme (choice of window size) and consequential border-misalignment, could also lead to more accurate segmentations.