A Combined Random Forests and Active Contour Model Approach for Fully Automatic Segmentation of the Left Atrium in Volumetric MRI

Figures

• 

  Figure 1. Flowchart of the developed algorithm.

• 

  Figure 2. A 2D illustration of the local and contextual features learning scheme. The red rectangle indicated a neighborhood B(x) around the center voxel x. (a) b = 0: Local features are captured from a scalable local region Rcentric centered at x, within the neighborhood B(x), and are fed into the contour evolution stage. (b) b = 1: Contextual features are captured from randomly displaced and randomly scaled local region Rrandom, and from Rcentric, within the neighborhood B(x), and are fed into the voxel classification stage.

• 

  Figure 3. Overview of the proposed concatenated classification framework. A single CRF is trained for an individual sample image. The labeling of a new target is performed by the testing step on each individual CRF and the fusion of the obtained probability maps.

• 

  Figure 4. The tissue probability maps estimated from individual CRF and multiconcatenated RFs for a target subject. (a) The axial slice from the original volumetric data. (b)–(d) Voxels with more than 0.6 confidence (green spots) in the outcomes of an individual CRF with iterations 1, 2, and 5, respectively. (e) Fusion of each individual CRF. (f) Ground truth.

• 

  Figure 5. Visual comparison of the proposed method with different components integrated on case B003 of the STACOM database. (a) Outcome of the volume-scalable ACM without shape constraint; (b) result of the CRFs without contour refinement (stage 2); (c) the combined CRFs and volume-scalable ACM (final stage); (d) ground truth.

• 

  Figure 6. Multiple slices of the 3D segmentation results for case B006 of the STACOM database. The final contours (shown as colored contours in the 1–7 columns) and the 3D surface models (shown as blue volumes in the last column) overlaid on the corresponding axial slices (a), the sagittal slices (b), and the coronal slices (c), respectively. The red contours are generated by the proposed algorithm, while the blue contours represent the ground truth, and the green contours are where the algorithm ones coincide with the ground truth.

• 

  Figure 7. Impact of 4 different parameters in the concatenated classification stage: the number of trees used per RF (a), the maximally allowed depth for each tree (b), the minimally allowed sample count per leaf node (c), and the iteration number of the concatenated scheme (d).