Modeling of abdominal perfusion on CT image sequences with contrast product injection

Scientific context: The solid tumors localized in the abdomen represent a major problem of public health. The new treatments against tumor angiogenesis (destruction of vessels which feed the tumor) are now part of the oncology therapy. Numerous clinical trials continue to show their usefulness by increasing the survival rate of many cancer affected patients. In these conditions, new functional imaging methods are being developed in an attempt to provide earlier markers of antiangiogenic treatment efficacy than the simple morphological criteria RECIST or his revised version RECIST 1.1 [Eisenhauer 09, Therasse 00]. Objective of the study: the main objective is to propose robust methods to allow estimation of functional markers reflecting the tumor evolution from dynamic studies. Latest developments: Currently, in this domain, assessing of the efficiency evaluation of an anti-angiogenic therapy is mainly based on the functional imaging of the microcirculation, which the objective is to quantify the tumor microcirculation by dynamic acquisitions with injection of contrast product. Despite numerous trials with bolus injection of contrast product, in CT, MRI or ultrasound [Eisenhauer 09, Marcus 09], the estimation of parameters is not reliable enough to be produced as additional criterion of evaluation in RECIST protocol. Issues: The current limitations are mainly due to a lack of control of the breathing movements, biasing studies of the time intensity curve on a pixel basis. Therefore, results are not homogeneous and may vary from a team to another one for the same treatment, particularly with CT imaging in which the axial sections are perpendicular to the breathing movements. In addition, no consensus exist on the parameters permitting the quantification of tumor microcirculation. Several approaches are based on a model of the diffusion of a contrast product in the biological tissues but different models have been developed. To the best of our knowledge, no rigorous comparison between these models exit and no recommendations are given to physicians to evaluate the conditions of application of these models. Finally, a step of parameters estimation of the selected models is performed. Due to the particular conditions of the CT acquisitions, this estimation must be computed form noisy data and a large number of regions (one estimation by voxel or group of voxel of the dynamic image sequence). Methods and results: The images of the sequence acquired in CT can be reconstructed in different ways (for example the thickness of the slices can be adjusted). A first study had the objective of determining the best values of these parameters in our context. To compensate the respiratory motion of the patient during all the CT dynamic acquisition, a registration method has been developed. This method is based on the multi-resolution block matching algorithm using a robust similarity measure to the diffusion of the contrast product during the acquisition. Two quantitative criteria have been proposed to evaluate the registration quality. Once the images are aligned, the functional bio-markers estimation is possible. To obtain a reliable parameter estimation, the first question concerns the choice of the considered model. A comparative analysis has been performed on five classical diffusion models that take the differential equations (ordinary or partial). This comparison is based on a rigorous mathematical framework (sensitivity analysis, detection of structural or practical non-identifiabilities, study of the cost function convexity) and gives a set of recommendations for the clinical use of these models. With this work, problems for the parameter estimation have been identified. An original parameter estimation method has been developed. This method, based on the 'multi-task' aspect, is a joint estimation for all image voxels. A new notion of similarity between tasks is introduced, taking advantage of similarities between data temporal evolution. The tests show us that this new similarity is better for the robustness of the estimation than the similarity based on the spatial neighborhoods. Conclusion: In a restrictive clinical context (noisy data, few number), we have developed a complete pipeline with a set of necessary steps to a reliable characterization of the tumor microcirculation from dynamic perfusion image sequence: the main contributions of this thesis cover the reconstruction parameters optimization, the development of a registration method, the argued selection of a perfusion model and the development of a robust method of parameter estimation. With these works, we can envision the utilization of these perfusion models to the characterization and the prediction of the therapy response of a patient. (author)