BME Frontiers / 2022 / Article / Fig 1

Research Article

Weakly- and Semisupervised Probabilistic Segmentation and Quantification of Reverberation Artifacts

Figure 1

(a) An example showing the difference in labels by different human annotators. The first image from the left displays the original image, whereas the rest are examples of different labels from different annotators on the same chicken breast image. The gray labels are needles and the white labels are reverberation artifacts. Annotators agree on the location of needles and reverberation artifacts, but disagree on the boundaries and the pixels between each reverberation. The second image from the left is an example of over-labeling, which is the labeling we use to label our training set. (b) Labels on the test set and segmentation results. The left image is a patch with needles and artifacts on a phantom image with the proposed test labeling overlayed on top, and the right two images are the segmentation results by our method of the needle and artifacts, respectively. In the left image, yellow and pink labeled areas represent the possible areas containing needles and artifacts, respectively, green and white labeled regions indicate the patches that are definitely the first and second reverberations, the bright blue circle(s) represent the non-artifact patches between reverberations, the dark blue region(s) are the most confident needle pixels, and the red patches are the regions with identifiable fuzzy artifacts. (c) Notation and abbreviations for the metrics.