Glandular tissue pattern analysis through multimodal mri-mammography registration

Resumen

Breast cancer is the most common cancer in women worldwide. Current statistics show that one in eight women will develop this disease over the course of her lifetime. Early detection increases the likelihood of overcoming the disease, motivating the implementation of screening programs. While X-ray mammography is the gold standard image modality for screening and diagnosis of breast cancer, it presents decreased sensitivity in dense breasts. Thus, other medical image modalities, such as magnetic resonance imaging (MRI) and ultrasounds, are used to overcome the limitations of the mammography. Several studies have shown that the combination of the different medical image modalities leads to a more accurate diagnosis and, therefore, a more effective medical treatment of patient diseases. However, the fusion of information among several image modalities is a challenging task, due to the differences not only in the physics underlying each modality but, also, the different patient positioning during the image acquisition. On the one hand, inner structures of the breast, such as nerves, blood vessels and ligaments, can be clearly visible in one modality but undetected in the others. On the other hand, movement restrictions, applied during the image acquisition, are completely different. The main purpose of this thesis is to evaluate the similarity among the information provided by two medical image modalities, such as the X-ray mammography and MRI, and, at the same time, to propose new algorithms to register the images in order to correlate the position of lesion and susceptible areas. To achieve this goal, we use registration algorithms and image analysis techniques. A deep review of the state-of-the-art, focusing our attention in the multimodal registration problems using patient-specific biomechanical finite-element (FE) models, is performed, from the biomechanical model construction (including the pre-processing and segmentation of MRI images, a suitable FE mesh construction as well as the methodology to quantify the accuracy and quality of the methods) to the physics underlying the mechanical deformation (elastic and hyperelastic parameters exposed in the literature, and loading forces and boundary condition) to solve the problem. Similarly, the registration algorithm and technical and medical aspects to validate the registration are exposed with the aim to bridge the gap between the engineering and the clinical performance. Finally, we include a brief state-of-the-art of software options as well as the requirements and advantages of each tool, in order to obtain a suitable simulation and an accurate solution of the problem. Our analysis begins evaluating the similarity of the glandular tissue between two mammograms from the same patient, acquired in the same day and in a short time frame. This fact allows us to evaluate the effect of the breast compression in the parenchymal pattern distribution. We use the commercial software Volpara$^{TM}$ in order to extract the glandular tissue distribution directly from full-field digital mammograms (FFDM). Afterwards, we analyze the breast volume, volume of glandular tissue and breast density values provided by the software. Regarding the density maps -i.e. the image of the glandular tissue distribution- we evaluate the information as well as the structural similarity by means of the image registration algorithms. Our results show that the information provided by Volpara$^{TM}$ is stable, independently of several acquisition factors, such as the X-ray energy spectrum or small variations in the angle of projection. However, changes in the compression -i.e. breast thickness- clearly affect the glandular tissue distribution in the image. The monomodal analysis provides us a baseline result to perform the multimodal comparison between MRI and mammography. To achieve this goal, a fully automatic framework to register Volpara$^{TM}$ Density Maps and MRI images was developed. This software uses a patient-specific biomechanical model of the breast, which mimics the mammographic compression performed during the mammographic acquisition. In this work, we propose a new methodology to project the glandular tissue directly from the MRI, avoiding the loss of information that can be yielded when the image is deformed. The evaluation of the similarity, between the real and synthetic density maps, also requires to propose a new methodology, which takes into consideration the multimodal nature of the problem. Measures and techniques from other fields such as information theory, visual perception and statistics are used in our work. Our analysis shows a high similarity between the information contained in the two modalities as well as a high structural similarity in the distribution of the glandular tissue. A visual analysis, using a checkerboard pattern, shows continuity in the longest structures of the glandular tissue. However, an aliasing effect is yielded when the glandular tissue is projected from the MRI During the registration framework evaluation we computed the target registration error (TRE) between landmarks -i.e. lesions- in both MRI and mammography. The 2D problem consists of directly projecting the landmark position from the MRI to the mammogram, computing the Euclidean distance between the computed and the real landmark position. However, locating the 3D position, within the MRI from the corresponding lesions in the mammograms, used to require complex and computational expensive methods to undo the breast compression. To solve this issue, we propose a new, fast and efficient algorithm to locate the landmark position within the MRI. Locating several rays from different mammographic projections (mainly, cranio caudal (CC) and mediolateral oblique (MLO)), the 3D landmark position can be computed as the intersection (or the center of mass of the two closest points) among the rays. Using a similar methodology to that proposed during the glandular tissue projection, a back-projection ray-tracing can be performed, allowing to locate the ray path in the uncompressed biomechanical model, avoiding to undo the breast compression. To accelerate the computation of the intersection between two rays, we propose an easy algorithm which sub-divide the rays, reducing the search space. Our model reduces the search space until 600 times with respect to a traditional point-by-point search. Thus, the computational time to locate the intersection is about $8~ms$, allowing real-time applications in the clinical practice. Furthermore, the TRE, in average, are about $1~cm$, better than those exposed in previous works. To conclude, we evaluate the capability of using other information, not only the glandular tissue (i.e. intensity), to perform the registration between the images. We focus our attention in the glandular tissue gradient. While the optimization algorithm (biomechanical model extraction and parameter optimization) is similar to our previous work, in this case the glandular tissue gradient is extracted from the mammograms and MRI images by means of image processing techniques. The MRI gradient is projected to the mammogram using two different approaches. The first one uses the accumulated directional derivative, obtaining a scalar value which is comparable to the norm of the gradients obtained from the mammograms. The second one accumulates the gradient considering each direction independently, yielding a vector which is comparable to the directional derivative obtained from the mammograms by means of a gradient correlation metric. In order to provide a fair comparison, we compare the results obtained using our methodology with a traditional intensity-based registration approach where we transform the MRI in a pseudo-Computer Tomography (pseudo-CT) image by means of a polyenergetic model instead of a traditional monoenergetic approach. Our results show an improvement in the TRE using scalar gradient values, with respect to the traditional intensity-based approach. Furthermore, we evaluate the different behavior using isotropic and anisotropic material models of the breast and we look for the correlation between the TRE and several factor of interest such as the breast glandularity and the inner landmark position. Anisotropic models shows an improvement with respect to the isotropic models, while the glandularity and the inner landmark position show a moderate correlation with respect to the TRE. Moreover, the computational time during the registration was reduced to half the time regarding our previous works. %, getting close to an acceptable value in the clinical practice. In summary, this thesis will help radiologists and physicists to better understand the variations of the glandular tissue that can be clearly visible in one modality but not in the other. Furthermore, the evaluation of the information can guide researchers to obtain more accurate segmentation algorithms, considering the partial volume effect presented in the MRI, as well as to improve the multimodal image registration between the two modalities, not only by means of intensity-based methods but also considering additional information such as gradients. Finally, our methodology includes several proposals to develop real time applications or with acceptable time values in the clinical practice.