A non-rigid registration approach for quantifying myocardial contraction in tagged MRI using generalized information measures

Abstract

We address the problem of quantitatively assessing myocardial function from tagged MRI sequences. We develop a two-step method comprising (i) a motion estimation step using a novel variational non-rigid registration technique based on generalized information measures, and (ii) a measurement step, yielding local and segmental deformation parameters over the whole myocardium. Experiments on healthy and pathological data demonstrate that this method delivers, within a reasonable computation time and in a fully unsupervised way, reliable measurements for normal subjects and quantitative pathology-specific information. Beyond cardiac MRI, this work redefines the foundations of variational non-rigid registration for information-theoretic similarity criteria with potential interest in multimodal medical imaging.

Multi-feature Statistical Nonrigid Registration Using High-Dimensional Generalized Information Measures

“…We used a local feature vector with normalized components, comprising the greylevel, and the eigenvalues and eigenvectors of the structure tensor ∂ xτ I(x) T ∂ xτ I(x) G σ of the image I, computed at a Gaussian scale σ equal to the half-width of the tagging pattern (D = 5). We studied the impact on registration accuracy of using a directional local contrast descriptor by comparing performances with a similar registration model exploiting solely the greylevel information [12]. To be consistent with the later work, we considered an infinite dimensional (non-parametric) transform space, and a registration criterion combining Havrda-Charvát information (α = 1.2) and the Nagel-Enkelmann oriented smoothness stabilizer, and performed experiments on the SPAMM dataset used in [12].…”

“…The proposed model has been applied to the estimation of myocardial deformations from tagged MRI exams by sequentially performing frame-to-frame registration [12]. We used a local feature vector with normalized components, comprising the greylevel, and the eigenvalues and eigenvectors of the structure tensor ∂ xτ I(x) T ∂ xτ I(x) G σ of the image I, computed at a Gaussian scale σ equal to the half-width of the tagging pattern (D = 5).…”

“…Whereas mutual information and its variations arise as a first choice [8], generalized information measures, such as f -informations [11,12] and Jensen f -divergences [9], have been shown to be relevant alternatives in specific clinical contexts.…”

Unbiased group-wise alignment by iterative central tendency estimations

“…For maximizing Equation (2.9), an iterative variational displacement field is derived similarly to the method described by Hermosillo et al [32] and Rougon et al [33]. These two works describe a framework for optimizing an image similarity functional S depending on a displacement field u.…”

“…As described in Hermosillo et al [32] and Rougon et al [33], the variational optimization of S is solved by looking for the displacement field that cancels the derivative in Equation (2.11). Such local minimum can be reached by iteratively composing the first variation given by Equation (2.18) with the current estimate of the displacement field.…”