AUTOMATED QUALITY CONTROL ANALYSIS OF DIGITAL MAMMOGRAPHY PHANTOM IMAGES

Abstract

Aims: Digital mammography (DM) screening has been widely adopted in the world for the last two decades; this has contributed to reducing the mortality rate due to early detection of breast cancer. Readers of mammography images need high quality mammography images to detect small lesions in the breast. However, the complexity of digital mammography systems requires rigorous monitoring and maintenance to ensure high image quality (IQ). This is achieved through periodic quality control (QC) as per mammography guidelines, with many adopting the use of the American College of Radiology (ACR) DM phantom. The IQ assessment though, remains subjective as it is still carried out by trained human readers. This might result in variations of assessment and some limitation in detecting small degradations of IQ. To overcome this challenge, this thesis aims to develop an automated analysis software for the ACR DM phantom. Method: The automated analysis of ACR DM phantom images was developed using a MATLAB application by applying different methods to analyse the images. For the scoring of the images, template matching with supplementary methods was mainly used for each target object (fibres, specks groups, and masses). Initially, the software was validated by comparison to three certified medical physicists observers using 27 images solely from one mammography manufacturer vendor. Those images were deliberately collected at different dose levels to ensure some variation in IQ. It was then further expanded to include 80 phantom images from several mammography manufacturers, which have different characteristics, and then we compared it with eleven human observers, including 6 experts and 5 non-experts. Lastly, we compared the scoring from ACR, and an in-house developed phantom called Applied Physics Group (APG). Furthermore, we examined the possibility of using processed phantom images for long term monitoring of mammography system performance. Results: The software scoring, in general, was comparable to the scoring by the human observers. In addition, the software was able to provide objective quantitative measurements such as contrast to noise ratio (CNR) and figure of merit (FOM) in an automated way. The comparison within one mammography manufacturer showed no significant differences in the scoring between the software and human observers. However, there was a significant difference in the scoring of low contrast objects (fibres and masses) when we included multiple mammography manufacturers and observers with different expertise. This could be due to the inter- and/or intra- variability of subjective assessment by human observers. Additionally, the findings show the potential of using processed images for purposes of long-term monitoring of mammography systems. Conclusion: The software has fully automated the quantitative analysis and scoring of mammography phantom images with minimal interaction by the users. The software would help the current quality assurance practice to provide consistent and objective outcomes to assess the IQ of mammography phantom images. This also would help mammography operators avoid subjectivity and visual acuity in the assessment of images.