Improving Segmentation of Breast Arterial Calcifications from Digital Mammography: Good Annotation Is All You Need

Kaier Wang, Melissa Hill, Seymour Knowles-Barley, Aristarkh Tikhonov, Lester Litchfield, James Christopher Bare; Proceedings of the Asian Conference on Computer Vision (ACCV) Workshops, 2022, pp. 130-146

Breast arterial calcifications (BACs) are frequently observed on screening mammography as calcified tracks along the course of an artery. These build-ups of calcium within the arterial wall may be associated with cardiovascular diseases (CVD). Accurate segmentation of BACs is a critical step in its quantification for the risk assessment of CVD but is challenging due to severely imbalanced positive/negative pixels and annotation quality, which is highly dependent on annotator's experience. In this study, we collected 6,573 raw tomosynthesis images where 95% had BACs in the initial pixel-wise annotation (performed by a third-party annotation company). The data were split with stratified sampling to 80% train, 10% validation and 10% test. Then we evaluated the performance of the deep learning models deeplabV3+ and Unet in segmenting BACs with varying training strategies such as different loss functions, encoders, image size and pre-processing methods. During the evaluation, large numbers of false positive labels were found in the annotations that significantly hindered the segmentation performance. Manual re-annotation of all images would be impossible owing to the required resources. Thus, we developed an automatic label correction algorithm based on BACs' morphology and physical properties. The algorithm was applied to training and validation labels to remove false positives. In comparison, we also manually re-annotated the test labels. The deep learning model re-trained on the algorithm-corrected labels resulted in a 29% improvement in the dice similarity score against the re-annotated test labels, suggesting that our label auto-correction algorithm is effective and that good annotations are important. Finally, we examined the drawbacks of an area-based segmentation metric, and proposed a length-based metric to assess the structural similarity between