Towards Precision Medicine in Breast Cancer Imaging : From 3D Breast CT Radiomics to 4D Perfusion
Breast cancer is the most diagnosed cancer in women, with an incidence rate, in the Netherlands, of one woman in seven (one in eight in the United States), if she lives to be 80 years old. To decrease its mortality and morbidity, multiple efforts have been undertaken in the past few decades to improve detection, diagnosis, and treatment of breast cancer. In all these applications, medical imaging plays or can play a key role. Several imaging modalities are currently being used in breast cancer care, with other modalities being investigated to further enhance detection, diagnosis, and treatment planning and monitoring. One of these modalities is dedicated breast CT.
Dedicated breast CT is a fully tomographic technique able to acquire three-dimensional (3D) images of the breast in a short time (10 to 16 seconds), with contrast, spatial resolution, and radiation dose characteristics optimized for imaging breast tissues and lesions. Given these properties, breast CT could be useful in multiple stages of the breast cancer imaging pipeline, potentially providing meaningful imaging biomarkers that may help characterize the biological signature of breast tumors. Automated analysis of these images to quantify imaging phenotypes can potentially result in better performance compared to the use of visual perception alone. This approach, called radiomics, aims at extracting several quantitative features from tumor images (e.g., describing texture, shape, and margin), usually after the tumor region has been identified and contoured by expert readers or dedicated software (a process called segmentation). After these quantitative features of the tumor are extracted, they can be used to develop mathematical and Artificial Intelligence (AI) models to aid radiologists in clinical decisions.
To investigate the power of such an approach in breast CT imaging, in the first part of this thesis radiomic algorithms were developed, validated, and applied to a dataset of mass images acquired with breast CT, with the aim of improving breast cancer diagnosis.
First, in Chapter 2, radiomic features able to quantify the tumor texture were implemented, and their stability across the segmentation results of multiple radiologists and an AI algorithm was studied. This resulted in both the development and validation of a segmentation algorithm for breast masses, and in the identification and elimination of those features that are strongly affected by the quality of the tumor contour delineation, thus being not suitable for radiomic analyses.
In Chapter 3, the breast CT dataset was enlarged thanks to the incorporation of a second dataset from another research site, and the previously developed texture features were supplemented with radiomic descriptors able to quantify different aspects of breast mass shape and margins, both established biomarkers of malignancy. To evaluate the power of these descriptors in the discrimination of benign and malignant cases, manual segmentation was used, and the stability analysis performed in the previous chapter was repeated for the new descriptors. This study resulted in an area under the receiver operating characteristic (ROC) curve (a metric varying between 0.5 and 1, reflecting the accuracy of diagnostic systems) in the discrimination of benign and malignant masses of 0.90.
In Chapter 4, a final radiomic model was developed, including the AI-based segmentation method and the radiomic descriptors developed in the previous chapters, combined with more advanced AI approaches (convolutional neural networks) able to automatically learn the relevant features directly from the images, without the need for segmentation or engineered extraction of radiomic features. Thanks to this hybrid computational approach, the area under the ROC curve was further improved, reaching a value of 0.95.
Despite the promising results of 3D breast CT radiomics for breast cancer diagnosis, this modality is still only anatomical in nature. To further improve its power in breast cancer characterization, acquisition of functional information would be crucial. This can be obtained by the use of contrast enhancing material: through the observation of its flow patterns within a tumor, information about vascular access and, potentially, cellular information could be obtained, leading to important biomarkers of biological signature. To achieve this, multiple acquisitions of the patient breast at different temporal moments should be performed after contrast medium administration, with spatial and temporal resolution being high enough to depict the enhancement characteristics of each tumor region.
Therefore, the second part of this thesis focuses on the development of a new modality, four-dimensional (4D) breast CT. Thanks to its spatio-temporal resolution properties, this modality is envisioned to achieve a higher level of breast cancer characterization compared to existing imaging modalities. Furthermore, if coupled with quantitative radiomic algorithms, it could open the path to a precision medicine approach where cancer treatment can be tailored to each single patient, based on specific tumor-related imaging biomarkers, potentially increasing treatment effectiveness and decreasing mortality and morbidity.
To develop and optimize this modality, tools are needed to model both the imaged object (the breast) and the imaging system, to simulate and optimize the image acquisition process in a virtual and safe environment, prior to imaging patients. To fulfil this goal, first, in Chapter 5, 3D patient breast CT images were segmented, resulting in discrete representations of the main tissue types found within the breast. This allowed for the generation of 3D breast phantoms able to model the anatomical composition of a patient breast.
In Chapter 6, information on the dynamics of blood flow in breast tissues and lesions were obtained from other literature, and these characteristics were added to the previously generated 3D phantoms, resulting in 4D breast models that mimic the functional behavior that may be found in this organ.
In Chapter 7, an AI-based method to increase the spatial resolution of the phantoms was developed. This process allows to overcome the initial resolution constraints of the phantoms (limited to the resolution of the 3D breast CT system used to acquire the image data), resulting in breast models able to better represent a real breast as a continuous object.
In Chapter 8, a 4D breast CT image simulation algorithm was developed and validated. This algorithm replicates, in a virtual environment, the process of imaging a breast while being perfused by iodinated contrast agent, resulting in simulated 4D breast CT images. This allows for the study of the effect of different image acquisition settings on the resulting image quality and, therefore, will help develop and optimize the first prototype of this imaging technique.