Introduction: Radiographers play a central role in the acquisition of high quality mammographic images, as image quality has a significant impact on diagnostic accuracy. To ensure high imaging standards within the national breast cancer screening programme DORA, image quality assessments of mammograms obtained by radiographers working in the programme are performed once or twice a year in Slovenia. This evaluation process is time-consuming and relatively subjective. Therefore, the use of automated tools to streamline the assessment procedures can both shorten the process and improve objectivity. Purpose: The aim of this dissertation is to test and develop various methods for automating the assessment of mammographic image quality within the DORA programme and to develop a dedicated web-based application that implements these methods. Methods: A cross-sectional study was conducted using an experimental approach. The study was conducted in several phases: Collection of image datasets and data acquisition; preprocessing of mammographic images; application of automation techniques (detection of key elements in the image, annotation of errors, breast segmentation methods, development and testing of convolutional neural networks [CNNs] to detect IMF errors, development and testing of object detection methods using the YOLO algorithm to detect keypoints, computational methods to determine objective quality metrics); and finally, the development of the web-based application. Results and discussion: The application development process took place between September 2020 and April 2024. We successfully developed and tested several computational approaches – from traditional segmentation methods to advanced deep learning models and object recognition systems. Thresholding proved to be sufficiently effective for basic breast segmentation, while CNNs and YOLO-based models achieved promising results in recognising and classifying individual quality criteria. Conclusion: The main outcome of this work is the development of a web application for evaluating the quality of mammographic images. When comparing the human evaluations with the application assessments, the best performance was obtained for criteria with clearly defined visual elements, while lower accuracy was observed for more subjective criteria and rare error types.
|
