Localized Probability of Mammographic Masking

Institution: University of California, San Francisco
Investigator(s): John Shepherd, Ph.D. -
Award Cycle: 2015 (Cycle 21) Grant #: 21IB-0130 Award: $124,112
Award Type: IDEA
Research Priorities
Detection, Prognosis and Treatment>Innovative Treatment Modalities: search for a cure

Initial Award Abstract (2015)

Interval cancers, diagnosed in the time after a normal screening examination and before the subsequent screen, are common in young women with dense breast, and as a general rule, have characteristics of rapid growth, and are frequently of advanced stage at the time of discovery. It has been reported that 68% of the 12-month interval cancer cases have high density breast. Importantly, this means that 32% of interval cancers are in women with low breast density. Thus breast density alone is not the entire picture. The quality of the density can obscure cancers as well. We call this quality texture. Furthermore, breast density and texture varies through all breasts. Describing the probability of a cancer hiding in a woman’s breast based on the average density and texture most certainly draws attention away from smaller regions of high masking probability, yet this is the only way masking has been describe to date. In this proposal, we create a method to report to the radiologist the local probability of masking throughout the breast. This approach personalizes the search for cancer by providing a means to target high masking regions of the breast with follow-up screening that removes the masking effect for that woman’s specific type of breast tissue. Having localized regions of high breast density with complex tissue texture in an otherwise lower density breasts is common. However, the frequency and size of these high masking areas are not known for women undergoing mammography. The studies that objectively assess the probability of local masking are lacking.

The question(s) or central hypotheses of the research: The central hypothesis of this proposal is that the water, lipid, and protein content of breast lesions varies sufficiently according to lesion type to provide clinically useful diagnostic information.

The general methodology: We plan to develop a method that can directly measure masking probability by superimposing cancer-like shapes into a woman’s mammogram and determine what is the minimum size and density of these objects that can be detected. Using this information, we can determine the probability for local regions in the mammogram masking a cancer. We will express this as a 0 to 100% detectability score. We will then calibrate this score in interval cancers in women and study what breast tissue qualities are then associated with a low detectability (high masking.) We will then review screening mammograms from 1000 women to estimate how common it is to have high masking regions of different sizes. Our group is uniquely qualified for this research since we have virtually all the software elements already developed – they just need to be used together. Lastly, when successful, this technique could be rapidly implemented since it doesn’t alter the way mammograms are acquired.

Innovative elements of the project: Our group is one of the few groups in the world that has validated an accurate and precise way to measure localized breast density. We also have developed a library of breast texture measures. A better understanding of the local tissue properties can be gained in terms of objective tissue quality measures such as localized volumetric breast density and breast texture. Once understood, adjuvant imaging modes that specifically address the particular masking effect for individual women could be used to target the high masking probability regions. Selectively probing of high masking areas may also reduce the time, cost, dose, and stress related to call backs.

Progress Report 1 (2016)

Detectability of invasive cancerous lesions in mammography is diminished in breasts with dense and complex fibroglandular tissue. If masking were locally quantified in mammograms, radiologists could clear the region using targeted screening techniques or other imaging methods. Our project quantified localized masking with an algorithm to determine detectability of virtual lesions inserted into the clinical mammogram.

Our project had three aims. First, we needed to develop automated software to create and insert virtual lesions based on the diameter and thickness of objects in the CDMAM phantom. We used X-ray attenuation physics to successfully create these virtual lesions and the attenuation of these images matched the attenuation of identical lesions in an actual CDMAM phantom. Our next aim was developing masking values and comparing how these values correlate with texture and density features. Thusfar we determined that many masking values correlated with BIRADS density and percent density, and we will compare these masking values to various texture features. We are also annotating mammograms to compare these masking values with incidence of interval cancer vs screen-detected cancers. Our last aim, determining the prevalence of these masking values in a cohort of women, will be accomplished in the future. During this project, we overcame several barriers and obstacles, including selecting the proper algorithm to determine detectability, determining proper ways to quantify the amount of masking, and increasing computational efficiency to make the algorithm run quickly.

Major accomplishments thusfar: creating virtual lesions based that match x-ray attenuation physics, inserting lesions and calculating detectability throughout the breast, optimizing code to speed up masking calculations, developing various global masking values, finding masking values correlate with BIRADS density score, getting images with interval and screen detected cancers annotated for future studies.

Plans for continuation of the project: Continue work associating density and texture with masking values; determining relationship between local and global masking values with cancer locations; determining relationship between masking values and incidence of interval vs screening-detected cancers; application of this work to tomography; and comparison of mammography masking values to tomography masking values.

Dynamic PET Image Reconstruction for Parametric Imaging Using the HYPR Kernel Method

Conference Abstract (2016)

A Measure of Regional Mammographic Masking Based on the CDMAM 3.4 Phantom

Benjamin Hinton1, Serghei Malkov1, Jesus Avila1, Bo Fan1, Bonnie Joe 1, Karla Kerlikowske2, Lin Ma2, Amir Pasha Mahmoudzadeh1, John Shepherd1
1 University of California-San Francisco – Department of Radiology
2 University of California-San Francisco – Department of Epidemiology and Biostatistics

Abstract: Interval cancers are defined as cancers diagnosed between normal screening mammogram intervals. One important category of interval cancers are cancers that were large enough to be detected at the time of screening but were masked by overlapping dense and spatially complex tissue. Masking diminishes detectability of tumors in women with dense breasts by 10-20%, which delays treatment and potentially increases cancer mortality. Further, interval cancers are often larger than screening detected cancers and of a more advanced stage once discovered.

We present a model of regional masking using an algorithm that determines the detectability of simulated lesions virtually inserted into raw mammography images. These lesions are based on the thicknesses and diameters of gold disks in the Contrast Detail Mammography (CDMAM) 3.4 phantom, which is widely used to determine contrast detail characteristics of mammography systems. We hypothesize that this model will produce regional detectability maps which would allow for clinicians to use other methods to clear regions with low mammographic detectability due to masking.

We first developed software to virtually insert these simulated gold disks and produced predictions of whether these gold disks would be detectable. We used training and validation data from a set of CDMAM images to tune our detectability algorithm and to validate these detectability thresholds that were produced. We found that for a given diameter of simulated inserted disk, our virtual detection algorithm predicted minimum detectabilities that were within the standard error of actual detectability measurements from the CDMAM phantom.

We then performed this calculation in 0.07 mm2 regions for the entire breast in a small selection of women who have had screening mammograms. We examined the Image Quality Factor (IQF), a measure of disk detectability where larger IQF values have more detectability and less masking, in regions of high breast density and low breast density. In regions with low breast density, the mean IQF value was 48.2 with a standard deviation of 4.7. In regions with higher breast density, the mean IQF value was 25.6 with a standard deviation of 4.6. This indicates that the IQF value differentiates between regions of high breast density and low breast density.

This is the first work to produce a regional map of mammographic masking based on the CDMAM phantom. This is valuable because mammographic masking is a key factor that reduces detectability of mammography and contributes to interval cancers. With a regional measure of masking, clinical radiologists could identify areas with low detectability and use other methods, such as ultrasound, to clear those regions with low mammographic detectability. This would help reduce the number of biopsies, the overall cost of breast cancer detection, and the stress and uncertainty associated with detecting breast cancer in dense breasts.