Modélisation 3D des artères hépatiques et visualisation par fusion d'imagerie par résonance magnétique et d'imagerie ultrasonore

Primary and secondary liver cancer constitute, in aggregate, the second most common cause of cancer-related death worldwide. The prognosis for these cancers is very poor with a 5 years survival rate under 20%. In Canada, the incidence rate of liver cancers has more than tripled over the past four decades, making it one of the fastest growing cancers in the country. Trans-arterial chemoembolization (TACE) is currently the standard of care for advanced cases of primary liver cancer. Intra-arterial therapies, such as TACE rely on the fact that blood supply to the tumor is predominantly of arterial origin while the portal vein is the primary blood supplier for the healthy liver parenchyma. Therefore, the hepatic arteries can be used to selectively target the tumor while sparing as much of the healthy liver as possible. During TACE, anti-cancer drugs are injected directly into the arteries feeding the tumor followed by the injection of an embolizing agent. The embolization prevents the drugs washout, thereby prolonging tumor exposure while also cutting of blood supply to the tumor, which provokes ischemia and tumor necrosis. Optimization eforts are underway to further maximize tumor exposure to the drugs while minimizing the toxic efects on healthy liver tissue. A possible strategy to optimize drug delivery would be to modulate the drug injection rate with regards to the real-time blood flow velocities in the feeding arteries. Such velocities can be obtained non-invasively using Doppler ultrasound (US). However, the hepatic arteries are only partially and sometimes not at all depicted in US images. Furthermore, abdominal ultrasound probes, which favor penetration depth over spatial resolution, are ill-fitted for the visualization of small hepatic vessels within the liver tissue. The numerous artefacts in US images (speckle noise, acoustic shadows) further complicate this task, especially for inexperienced users. Therefore, the acquisition of the blood flow measurements on these vessels is challenging. In order to overcome this difculty, we propose to combine a patient specific 3D model of the hepatic arteries with a 3D US image. After the fusion, the user can revert back to 2D US images, which will contain segments of the 3D model, thereby highlighting the localization of the arteries. Building a tool that allows for the integration of a 3D model of the hepatic arteries in the US image is the main objective of this master's project. We generate the 3D model of the hepatic arteries from 3D arterial magnetic resonance an¬giography (MRA) data using an in-house algorithm based on the propagation of tubular deformable models along the vessel centerlines detected by a multi-hypothesis tracking mod¬ule. Combining this model with the US image requires registering the arterial MRA, from which the model is extracted, with the US. This is a difcult problem mainly because the two images have very diferent aspects. The direct registration is further complicated by the scarcity of common features between the two images. Therefore, we chose to implement a compositional registration strategy. We use the portal phase of the MRA as an intermediary between arterial MRA and US. Portal veins are larger vessels than can be easily visualized in both modalities and can serve as common features to guide the MRA-US registration. We start by co-aligning the two phases of the MRA using a non-rigid registration to account for possible liver deformation between the two images. The portal phase MRA is then non-rigidly registered with the 3D US using a 3 step method. First, a rigid alignment is performed between the two images by matching masks of the portal and hepatic veins extracted from both modalities. Second, an afne registration is performed using an intensity-based similarity metric that postulates the existence of a local linear combination between the US image intensities and both the intensities and gradients of the portal MRA. We also incorporate a mask of the liver, segmented from the portal MRA, in the computation of the metric in order to limit the contribution of non-liver voxels to the value of the metric. Finally, a deformable registration using the same metric is applied to account for liver motions that cannot be completely corrected by rigid or afne transformations only. Finally, the solution to the arterial MRA-portal registration problem is obtained by com¬posing the two previously determined transformations (i.e the transformation from arterial MRA to portal MRA and then the transformation from portal MRA to 3D US) into one global mapping function between arterial MRA and US. This approach assumes that, since the hepatic arteries are known for anatomically accompanying the portal vein, the transfor¬mation between portal MRA and 3D US can be used to register the arterial MRA and US images, after having previously registered the two phases of the MRA. The validation of our tool on a dataset of 10 porcine models yielded promising results. The registration accuracies were measured in terms of target registration error (TRE) between pairs of corresponding landmarks selected in both modalities before registration. The mono-modal registration between the MRA images yielded a mean TRE of 1.2 ± 0.6 mm. The mean TRE values for the multi-modal MRA-US registration were 8.6 ± 3.0 mm for the rigid alignment, 4.0 ± 1.0 mm for the afne registration and 3.9 ± 1.2 mm for the deformable reg¬istration. Statistical analysis showed that both afne and deformable registrations provided a significant improvement (p ≠ value < 0.002 in both cases) from the initial rigid alignment. However, the diferences in TRE between afne and deformable registration were not statis-tically significant (p ≠ value =0.734). We also studied the robustness of our method to the quality of the registration initialization by performing 100 randomly initialized registration experiments for each pig. The initial rigid parameters were randomly perturbed by adding white Gaussian noise (standard deviation: 10 mm or 10¶) to each of those parameters. We reached a mean convergence rate of 69.6 ± 13.9 mm. Our validation also evaluated the efect of removing the liver mask from the computation of the similarity metric on the registration results. We showed that the mask prevents the occurrence of misregistrations that occur when the liver edges in the US image are matched with strong edges outside the liver in the MRA image. Finally, we presented an extension of the validation of the segmentation algorithm by comparing the meshes automatically generated from the MRA images of 15 hu¬man patients with the meshes manually segmented from computed tomography angiography (CTA) images acquired on the same patients. This comparison is based on the evaluation of three metrics: the mean absolute distance, the mean squared distance and the Hausdorf distance. We obtained the following results when comparing the MRA mesh to the CTA gold standard: MAD:1.05 ± 0.35 mm, MSD:2.38 ± 1.93 mm and HD:6.69 ± 3.45 mm.