Multicontrast 3D automated segmentation of cardiovascular images

Our work explores the use of multiple contrasts to perform automated 3D cardiovascular image segmentation, with the goal of facilitating 3D assessment of subject-specific heart models.

The signal intensity of MR images is a function of intrinsic tissue parameters (e.g., spin density ρ, T₁, and T₂), and chosen sequence parameters (e.g., flip angle, T_R, T_I). MR images do not directly measure the underlying states of the tissue parameters, but rather provide observations of functions of the underlying states. When sequence parameters are selected properly, the contrast between tissue types is high enough to allow differentation of tissues via image segmentation.

Our work explores image segmentation using multiple image contrasts (i.e., multiple acquisitions using different sequence parameters). With multiple contrasts, the signal intensities from different images can be used to form a multidimensional feature space, increasing the distance between observations of different tissues as compared to a one-dimensional feature space based on a single image contrast.

A proof of concept 3D IR-bSSFP scan was collected with different inversion times; all other parameters were the same for both scans using T_I = 49 ms and T_I = 330 ms. Images were processed using the FMRIB Software Library (FSL). Images were registered with the FLIRT tool, employing mutual information as the similarity metric, then cropped to include only the heart. The cropped, registered images were segmented by inputting both contrasts into the FAST tool. The number of tissues was specified as four, with the intent of differentiating air, fat, blood, and myocardium.

Figure 2 shows segmentation results using: left, manual drawing (for reference); center, FAST with one image contrast (T_I = 49 ms); and right, FAST with both image contrasts (T_I = 49 ms and T_I = 330 ms). Using only one contrast results in large areas of tissue misclassification, as indicated by the arrows. The use of multiple image contrasts results in vastly improved tissue classification.

As more advanced 3D viewing tools become available, the need to automate the segmentation process grows. This simple technique demonstrates a shift in imaging goal from high resolution 2D acquisition to tissue identification and lays the foundation for future sequence design.