MRI-based breast density estimation

Breast cancer is the most commonly diagnosed cancer among women, bringing significant burden to patients and society. On average, 1 in 8 women will develop breast cancer over their lifetime in the US. Breast density (BD) has been established as an independent and significant risk factor for breast cancer with higher BD conferring a 3-5 fold higher lifetime risk of breast cancer. BD is a radiologic (image-based) measure of the proportion of fat to fibroglandular tissues in the breast. Dense breasts have more fibroglandular tissue and less fatty tissue, indicating a higher risk for breast cancer. BD has emerged as a potential modifiable risk biomarker with change in BD increasingly incorporated as an intermediate surrogate endpoint in clinical trials to evaluate efficacy of drugs for the treatment and prevention of breast cancer.

As such, accurate breast density estimation becomes as a priority to detect small BD changes for assessing the breast cancer risk. Currently, mammography is the most widely used method for BD determination (MG-BD) in clinical practice. MG-BD is obtained by exposing a compressed breast to low dose x-rays, which leads to several limitations: 1) ionizing radiation prohibits its use in studies requiring frequent monitoring; 2) breast compression causes patient discomfort and limits accuracy due to the overlapping of breast tissues; 3) the reliability of MG-BD is also relatively low and wide because a slight difference in x-ray exposure calibration can lead to different MG-BD. All of these factors suggest that MG-BD may not be accurate and sensitive to small changes in BD.

MRI provides a safe alternative for BD quantification in the entire breast volume without ionizing radiation; this is particularly important for frequent monitoring as employed in clinical trials and for high risk patients. Fat-water decomposition MRI is a particularly useful technique for BD estimation, which is capable of representing the true breast tissue composition by separating MRI signal coming from protons in water molecules and in fat molecules using the Dixon method.

Utilizing fat-water decomposition MRI, we developed a fully automated and highly reproducible MRI-based BD measurement (MRI-BD), which is free of ionizing radiation and directly comparable to mammographic density (MG-BD).

We acquired MR images on both GE and Siemens scanners. On the GE scanner, radial IDEAL-GRASE sequence was performed, while on the Siemens scanner, a 3D Cartesian 6-echo gradient echo pulse sequence was conducted. IDEAL fat-water separation technique was performed: fat-only images and water-only images can be fast reconstructed from multi-echo data by applying a signal model to the phase evolution at different echo times. A validated automated breast segmentation was then applied. BD can be then quantified by FraGW(=FraGland+FraWater), which accounts for the volumetric fraction of the fibroglandular tissue and the actual water content in the breast after mathematically correcting the fat-water signal bias.

FraGW has been shown to strongly correlated with MG-BD (Spearman correlation coefficient = 0.96, p < 0.0001). Based on this correlation, a calibration curve has been established to convert FraGW to the MRI-BD. Our preliminary data demonstrate that MRI-BD can achieve extremely high reliability, with minimal test-retest variations (1.1 +_ 1.2%) and high intra-class correlation coefficient (0.99). This surpasses the intra-reader correlation coefficient of mammographic density (0.92) in our quality control procedure. This highly reproducible MRI-BD measure is directly comparable to MG-BD and enables the early detection of small BD changes, offering an important tool for assessing breast cancer risk and for evaluating the effects of prevention strategies aimed at reducing BD.

MRI Radiomics in predicting sentinel lymph node metastasis in breast cancer

Metastasis is the leading cause of mortality in breast cancer patients. Sentinel lymph node (SLN) is the first draining site to be affected during tumor spreading, and thus SLN status offers a valuable prognostic factor to guide treatment decisions. It is the general consensus that a complete axillary lymph node (ALN) dissection (ALND) is not necessary for patients without SLN metastasis, and thus the associated serious complications can be avoided. Therefore, non-invasive assessment of SLN status is extremely important: for patients with negative SLN, no lymph nodes need to be removed (hence avoiding the unnecessary, significant potential complications), while for patients with positive SLN, treatments including SLN dissection, ALND or radiation therapy should be further evaluated by clinicians depending on the tumor's characteristics.

However, SLN status is currently clinically determined by SLN dissection, also known as SLN biopsy (SLNB), which suffers from several limiations. It is an invasive procedure linked to significant complications, including: shoulder dysfunction, nerve damage, lymphedema and upper arm numbness. In addition, SLNB has been considered a controversial procedure, owing to the unstandardized radiopharmaceuticals, injection sites, experience of the operator, etc. Furthermore, SLNB is associated with a high false-negative rate ranging from 5.5% to 16.7%. Therefore, non-invasive methods are highly desirable to preoperatively evaluate SLN metastasis.

The existing non-invasive prediction methods of SLN status based on clinicopathologic characteristics of the primary tumor (such as MSKCC nomogram) have yet to show satisfactory results. The values of the area under the receiver operating characteristic (ROC) curve (AUC) of these methods were all lower than 0.8, which is insufficient to reliably guide clinical practice.

The increased use of MRI in routine clinical evaluation of breast cancer enables the possibility of non-invasively predicting SLN status based on the radiomic information of the primary tumor. MRI does not employ any ionizing radiation, and has been widely used in breast cancer detection and staging. Radiomics, an emerging technique of extracting high-dimensional quantitative image features to provide potential biomarkers for guiding clinical decisions, has drawn increased interest with recent advancements in pattern recognition tools. Radiomics-predicated techniques offer the possibility of establishing a completely non-invasive predictive model of SLN metastasis from the existing MRI data.

We performed a retrospective study on 212 breast cancer patients (69 positive SLN and 143 negative SLN) with DCE-MRI images. The intratumoral regions-of-interest (ROIs) were manually drawn by an experienced radiologist and peritumoral ROIs were obtained by dilation. A total of 590 radiomic features (including shape, histogram, texture and Laws features) were extracted for each patient from both intratumoral and peritumoral ROIs in the wash-in maps ((S1-S0)/S0)x100%, wash-out maps ((S1-S4)/S1))x100% and signal enhancement ratio (SER) maps ((S1-S0)/(S4-S0))x100%, where S0, S1, and S4 are the pre-contrast, first post-contrast and fourth (the last) post-contrast images, respectively.

The dataset was then randomly separated into two sets: a training set (~67%) and a validation set (~33%). The prediction model was built using the training set by logistic regression with the most significant radiomic features after feature selection combined with clinicopathologic characteristics. The prediction model was further evaluated in the independent validation set. Seven radiomic features were automatically selected from the training set and combined with the 7 clinicopathologic characteristics to establish a logistic regression prediction model of SLN metastasis. In the training set, AUC=0.961, sensitivity=0.880, specificity=0.948 and negative predictive value (NPV)=0.892. In the independent valuation set, AUC=0.866, sensitivity=0.783, specificity=0.872 and NPV=0.891.

This is the first attempt to combine radiomic analysis of routine DCE-MRI and clinicopathologic characteristics of the primary tumor to predict SLN metastasis for breast cancer patients, and the non-invasive prediction model achieved a high AUC and NPV in the independent validation set. This method can greatly benefit patients with invasive breast cancer, especially for those with negative SLN, helping them avoid unnecessary invasive lymph node removal and the associated serious complications. This study is a step further towards precision medicine and personalized treatment for breast cancer patients.

Why quantitative MRI?

Conventionally, magnetic resonance imaging (MRI) is used in a qualitative manner. When an image is obtained using the MRI scanner, radiologists look at the images with varies contrasts (Proton Density, T1 weighted, T2 weighted, etc) and identify regions with abnormal contrast to naked eye. For example, a region showing higher contrast in a T2-weighted image will be reported as hyperintense on a T2w image.

The story of quantitative MRI dates back to 1971, it was discovered that some tumors have higher relaxation times than regular tissue. While some tissue changes can be easily identified on a conventional MRI, other tissue with subtle T1 or T2 changes can not be picked up by naked eye. These changes can be identified with quantitative MRI.

Quantitative MRI offers several advantages including higher sensititive, better accuracy and higher reproducibility.

If quantitative MRI is this great, what is the reason it is not used every day in clinical MRI (yet)? The main problem is the significant longer acquisition time to obtain the quantitative maps (T1 map, T2 map, etc). For example, the conventional way to obtain a T2 map requires the acquisition of 4-32 T2w images. This effectly increase the acquisition time by 4-32 folds which is not acceptable in a clinical setting. This is where my research comes in.