Radiomics of breast cancer in magnetic resonance imaging
Loading...
Date
Authors
Μανωλίτσης, Αντώνιος-Μάξιμος
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Up-to-date large-scale analysis has designated breast cancer as the most prominent cancer type in terms of incidence and mortality rates in female population worldwide. More specifically, female breast cancer incidence rates remain relatively stable at around 24.6% over the last 15 years, while mortality rates show a declining tendency and currently stand at 15% showcasing the progress made in this area.
Given the extend of the problem, a rigorous analysis of the underlying complexity (heterogeneity) of breast lesions might prove of vital importance in addressing this issue, introducing a new generation of personalized/targeted medicine. Tumour heterogeneity is generally classified either as inter-lesion or intra-lesion and so far, is principally assessed using specific genotyping analysis that incorporates biopsy-proven biomarkers. However, the invasive nature of these methods creates the demand for an alternative standardized and quantitative evaluation of breast intra-lesion heterogeneity based on its phenotypic expression captured by imaging modalities.
Apart from mammography which still remains the golden standard in tumor diagnosis, MRI has also played a central role in breast cancer diagnosis being currently considered as the most sensitive diagnostic modality in breast imaging. The fact that MRI is a non-ionizing modality that provides a remarkable soft tissue contrast makes it ideal for specific cases such as women with dense breast tissue, which is the reason that it is suggested as an optimal supplementary screening modality. The rapid development of imaging technology in the last years has introduced a series of multi-parametric features of magnetic resonance imaging (MRI) which have lead in a more accurate breast cancer detection and characterization.
Two of the most important ones are Dynamic Contrast Enhanced (DCE) and Diffusion Weighted Imaging (DWI).
DCE-MRI is presumably the best modality for exploiting angiogenesis and is generally considered as an optimal means of assessing the local extent of the tumor and tumor heterogeneity. As implied by its name, it examines the temporal enhancement pattern of a tissue given the administration of a paramagnetic contrast agent into the vascular system building the basis for the extraction of several quantitative or qualitative parameters regarding tissue perfusion and micro-vascular status. In addition, it is highly capable of identifying tumour contours based on vascular density and permeability leading to an improvement of sensitivity in cancer detection between 86% to 99%. Regarding breast cancer, DCE has been extensively utilized for detection, diagnosis, characterization of breast cancer, and prediction of treatment response, mainly due to the high spatial resolution it offers. In fact, DCE-MRI has been recognised as a standard diagnostic technique included into the Breast Imaging Reporting and Data System (BI-RADS) that is defined by high sensitivity and variable specificity (from about 67% to 72%) in terms of breast lesions characterization.
DWI on the other hand, is based on the diffusivity of water molecules in tissues of interest. Variations in the diffusion properties of these molecules among different voxels are captured through motion-probing gradients leading to the formation of an image contrast characteristic of the tumour physiology. Many different approaches attempt to interpret this image contrast by introducing quantitative and qualitative measures, able to assess the tumour cellularity and integrity of cell membranes. The most common and easy approach of quantifying DWI is by computing for each respective image voxel its corresponding apparent diffusion coefficient (ADC) value. This value defines the average area covered by a molecule per unit time (mm^2/sec) reflecting in a specific way the magnitude of water diffusivity in that region.
Current literature has provided substantial evidence regarding the significance of ADC measurements in breast tumour classification revealing an inverse correlation between tumour cellularity and ADC values. ADC values have also demonstrated a strong connection with prognostic biomarkers such as subtypes or hormone receptor status, while their relative change between examinations has been proven capable of assisting in the evaluation of clinical response or treatment outcome. Furthermore, many recent studies have showcased the potential of DWI to act as a complimentary modality to DCE-MRI in order to enhance the overall diagnostic accuracy and specifically the poor specificity characterizing breast MRI. The main issue with DWI is that ADC values depend on several factors (b-value, diffusion time, imaging protocol, scanner) which is translated as a variability in ADC value thresholds making essential the introduction of a standardization scheme in order to be able to obtain clinically applicable biomarkers.
One of the main reasons behind the wide success of both of these approaches (DCE and DWI) in assessing breast tumour heterogeneity is the introduction of radiomics. In a nutshell, radiomics can be described as the extraction of multiple quantitative image features based on a ROI, followed by their extensive analysis aimed on decision making. The main target of radiomics is to provide accurate and robust biomarkers able to reveal the underlying tumour heterogeneity and quantify intrinsic tumour properties as expressed by the imaging modality. This is achieved through the extraction of texture-based features able to quantify local spatial distribution variations of gray levels in the image based on either a statistical or structural approach.
In the case of breast cancer, statistical approaches are generally highly exploited providing encouraging results. In brief, statistical methods are separated into three categories according to the number of pixels that take part in the formation of the features. First-order statistics (histogram-based) describe only the distribution of grey levels in an image, while second-order express the spatial relations between pairs of pixels (GLCM,GLRLM,GLSZM,NGTDM). Higher-order statistics, which are also called filter-based, apply filter grids on the image to reveal repetitive or non-repetitive patterns (Gabor filter, wavelets, Laws features). Given its high-dimensional nature, radiomics require robust analytic tools in order to achieve their full potential. These tools are provided in the form of machine learning algorithms which lead in the construction of powerful models that can eventually contribute actively in decision support.
Regarding DWI and especially ADC maps, a significant amount of studies have exploited several statistical-based textural features extracted from the whole region of interest which demonstrated a relative efficiency regarding diagnostic, prognostic, and predictive accuracy for response to treatment and survival outcomes.
This thesis, however, focuses on a different approach of feature extraction from ADC maps based on a nxn sliding window inside the ROI and not on the whole region. More specifically, distinct feature values are assigned to each respective pixel of the region of interest computed based on either a 3x3 or a 5x5 window around it. This overall leads to the construction of radiomic feature maps (RFM's) which we are going to evaluate in terms of their ability to accurately distinguish benign from malignant breast tumours. The basic reason we adopted this approach is because it restricts the feature calculation in a local window which might potentially make our final features more sensitive to local-based heterogeneity.
While this approach has already been applied in DCE-MRI for several occasions (e.g. for prediction of prostate cancer recurrence, for prediction of event free survival in anal cancer treated by chemoradiotherapy and particularly for breast cancer for molecular subtype classification, prediction of response the chemotherapy and benign and malignant tumour classification) mainly due to the higher spatial resolution it offers, with the exception of one no analogous efforts have been attempted to our knowledge based on ADC maps. This fact makes it quite intriguing to try comparing the performance of radiomic feature maps constructed upon ADC maps with the respective performance of regular whole-region based radiomics approach and discover if those maps can actually offer additional value except from visualization of the tumor.
Description
Keywords
Machine learning, Radiomics