3DQuorum™ Imaging Technology

Introduction

The use of breast tomosynthesis in screening has demonstrated an improvement in cancer detection
and a reduction in false positives compared to digital mammography.1 Some of the challenges of the routine use of breast tomosynthesis have been the increase in the number of images that need to be reviewed by radiologists and an increase in the average interpretation time.2 The large file sizes also require adequate PACS storage and high network bandwidth. The recent introduction of high resolution tomosynthesis imaging with 70-micron pixel sizes (Hologic 3Dimensions®) further affects these issues.

In response to these challenges, Hologic has introduced a new type of tomosynthesis images, called 3DQuorumTM, which provides 6-mm thick slices (known as SmartSlices) with 70-micron pixel resolution and utilizes artificial intelligence and machine learning, which along with the Intelligent 2DTM synthesized 2D images, provides the same diagnostic performance as 1-mm thick slices and with a faster average interpretation time as demonstrated in a clinical trial presented to the US FDA.

Design Goals

The most commonly used slice thickness in commercial breast tomosynthesis systems is 1 mm.3 For a common breast thickness of 60 mm, the radiologist needs to review 240 image slices in a standard 4-view screening mammogram set. The design goal of 3DQuorum is to provide fewer, thicker slices for review which speeds up interpretation time, uses smaller files for ease of transfer and storage, and at the same time maintains clinical performance compared to 1-mm breast tomosynthesis datasets.

Some commercial systems provide 10-mm slices, called slabs.4 However, these 10-mm slabs are presented in addition to the thin 1-mm slices, and so they both increase file sizes and are not likely to decrease radiologist interpretation time as they must be reviewed in addition to the 1-mm slices.

It is not readily apparent what the optimal slice thickness should be in breast tomosynthesis. One could make the argument that 1-mm default slices are thinner than needed, given the clinical need of detecting lesions of size 5-10 mm, and the detection of microcalcification clusters where the calcifications are distributed in the z-direction by 10 or more mm. In the case of a 10-mm lesion with distortionsor spiculations pointing out of the plane, it may be better appreciated when the entire lesion can be seen in focus in a given slice, which will only happen with a slice thickness greater than 1 mm.

Using microcalcification clusters as another example, seeing all the calcifications that form a cluster in one slice may aid detection, and can argue for a slice thickness greater than 1 mm. Additionally, searching for microcalcification clusters in hundreds of slices is likely to be slower and more tiring than searching for them in a smaller number of slices.

Design Goals
- Faster radiologist reading time
- Reduced number of tomosynthesis slices
- Reduced data storage space and network traffic

With these observations in mind, Hologic investigated what was an optimal slice thickness greater than 1 mm which could replace the 1-mm slices. While the first design criterion was to ensure maintaining clinical performance, during very initial experiments it was clear that it was necessary to maintain an overlap between adjacent thicker slices for smoother transition during scrolling. This would also ensure that no object was sub-optimally displayed if it happened to lay between two slices. An investigation was conducted to evaluate the performance of a range of slice thicknesses for the SmartSlices, from 4-mm slice thickness with 2-mm overlap, up to 10-mm slice thickness with 5-mm overlap. 2-mm SmartSlices with 1-mm overlap offers no reduction in the total number of slices so that option was rejected in advance. In order to determine optimal slice thickness and overlap size, a series of side-by- side image reviews were conducted where 2 to 3 users with a range of experience in reading tomosynthesis images carefully evaluated lesion appearance of a variety of cancerous and benign lesions. Starting with the configuration of 10 slices with 5-mm overlap, side-by-side evaluation included 8 mm thickness with 4mm overlap, 6 mm thickness with 3 mm overlap and 4 mm thickness with 2 mm overlap. The cases used for this subjective evaluation incorporated several lesions with small size (less than 1 cm) and low conspicuity to understand if any clinical information is lost during the process of creating SmartSlices. The users were also asked to evaluate appearance of normal tissue structure and report any potential additional call back that could be attributed to the process of creating SmartSlices. The users expressed that 10 mm slice thickness with 5 mm overlap and 8 mm slice thickness with 4 mm tend to show marginal loss of conspicuity when compared with 1 mm reconstructed images, especially for very subtle lesions that consisted of amorphous microcalcification clusters or small architectural distortions in dense breasts. The remaining configurations (6 mm and 4 mm) were deemed to show no loss of conspicuity on any of the lesions evaluated, and also did not indicate any concerns of false recalls due to the creating of SmartSlices. Since 6 mm thickness with 3 mm overlap was more advantageous than 4 mm thickness in reducing the number of images, that option was chosen for the final product. In addition to this subjective evaluation a pilot study with MRMC study design was conducted on approximately 100 cases and 5 readers to understand the difference in AUC using the 1 mm slices and the proposed configuration of SmartSlices at 6 mm. In this pilot study the AUC of the two reconstruction modes was comparable, with the 6 mm SmartSlice mode performing marginally better than 1 mm tomosynthesis mode for AUC. This confirmed that the selected configuration did not result in any loss of clinical information while reducing the number of images to one third of the original 1 mm slices.