Automatic Cropping for LV Segmentation in Cardiac MRI

Abstract

Left ventricle (LV) contour extraction and tracking is important in diagnosis of cardiac pathologies. This procedure is difficult, is usually performed manually using software assistance, is time consuming and presents high intra/inter operator variability. Automatization of this process is of great importance to the clinical practice.
The most state of the art algorithms for LV segmentation, described in the literature,are semi-automatic allowing to reduce significantly the human intervention. A full automatic approach is still an open problem mainly because it is difficult to compute a good initial guess for the contour, close enough of the final solution. Additionally, since the total amount of available data is usually large, and the region of interest (ROI) containing the LV is smaller than the overall acquired volume, an automatic crop is needed to reduce the computational burden associated with the pre-processing and segmentation algorithms. The pre-processing performs compensation in one hand, for coil sensitivity that induces intensity fluctuations across the images and, in the other hand, for misalignments due small movements during the breath-hold acquisition protocol.
In this paper we propose an algorithm to automatically crop the acquired volume by defining an adaptive size ROI containing the LV. This cropping is based on the well known Hough transform for circles, in the prior knowledge about the location of the heart and also on the high signal variability over time due to muscle contraction at the LV borders.Tests with real data have shown promising results.