Keywords:
Artificial Intelligence, Breast, Radiographers, Mammography, Screening, Technology assessment, Education and training, Image verification, Quality assurance
Authors:
H.-M. Gilroy1, M. L. Hill1, A. Chan1, M. Halling-Brown2, R. P. Highnam1; 1Wellington/NZ, 2Guildford/UK
DOI:
10.26044/ecr2021/C-13781
Methods and materials
A dataset of four-view mammographic screening studies was retrospectively identified, as per Figure 1
, from the OPTIMAM image database (screening mammograms from two sites in the UK NHSBSP between 2011-2018) among women without findings (“normal”) for two or more screening rounds.[5]
There were 2,134 “technical repeat” (TP) studies identified which included a single, same-day repeated view and reason codes related to inadequate positioning. [6] There were 1,275 images with repeat code R1A, i.e., “Inadequate positioning – Client” and 859 images with repeat code R1B, i.e., “Inadequate positioning- Radiographer.”
A second group of 1,340 “accepted” (no repeated views) studies were also identified for comparison.
Volpara Density and Volpara PGMI (Volpara Health Technologies Ltd., NZ) were used to evaluate breast density and image-level breast positioning. The software uses a series of algorithms to segment and identify key landmarks in the breast, and then carry out several breast measurements using the identified landmarks before assessing the positioning related to the following component metrics:
- Whether or not the nipple was in profile
- If tissue has been cut off
- The difference in posterior nipple line (PNL) between CC and MLO of breast side
- The adequacy of the pectoral muscle visualised (MLO only)
- If the pectoral muscle reaches the PNL (MLO only)
- The pectoral shape (MLO only)
- The presence of pectoral skinfolds (MLO only)
- The visibility of the infra-mammary fold (MLO only)
- Whether the nipple was pointing medially or laterally (CC only)
The software then assigns image-level and study-level grades of Perfect, Good, Moderate, or Inadequate (PGMI) based on a combination of the above metrics.
Univariate statistical analyses were used to investigate the potential associations between Volpara PGMI, TP and accepted image features. Chi-square tests were performed for the component metrics and Wilcoxon-Mann-Whitney tests were used for analysis of population distribution. All statistical analysis was performed using Stata v13.0.