Fast tissue classification in dynamic contrast enhanced

A contrast enhanced dynamic Magnetic Resonance clinical exam produces a set of images displaying over time the concentration of a contrast agent in the blood stream of an organ. The portion of tissue represented by each pixel can be classified as normal, benign or malignant tumoral, according to the qualitative behavior of the contrast agent uptake associated to it. These responses can be considered as the noisy output of a pharmacokinetic distributed model whose parameters have an intrinsic diagnostic importance. Fundamental MR imaging characteristics force a compromise between the noise level and the spatial and temporal resolution of the dynamic sequence. This makes the identification of the pharmacokinetic parameters and the classification problem difficult especially if short computation time is required by physicians. In this paper, a fast method is proposed to solve simultaneously the parameter identification and the classification problems. The complexity of the algorithm is $O(N\cdot n_p)$ flops where $N$ is the number of pixels and $n_p$ is the number of pharmacokinetic parameters per pixel. A family of functions for the parameters and the classification labels is defined. Each function is the weighted sum, with unknown weights, of a coherence-to-data term, several terms which enforce a roughness penalty on the model parameters, a term measuring the distance between the parameters in each pixel and the expected parameters for each class and a term which enforces a roughness penalty on the classification labels. A constrained optimization problem is solved to choose a member of the family, i.e. to estimate the unknown weights, and to minimize it in order to jointly estimate the parameters and the classification labels. A tuning procedure have been also devised, which makes the algorithm fully automated. The performances of the method are illustrated on real data sets.