Development of Multiscale 3D Residual U-Net to Segment Edematous Adipose Tissue by Leveraging Annotations from Non-Edematous Adipose Tissue

Improved subcutaneous edema segmentation on abdominal CT using a generated adipose tissue density prior

PURPOSE

Edema, or swelling, is a common symptom of kidney, heart, and liver disease. Volumetric edema measurement is potentially clinically useful. Edema can occur in various tissues. This work focuses on segmentation and volume measurement of one common site, subcutaneous adipose tissue.

METHODS

The density distributions of edema and subcutaneous adipose tissue are represented as a two-class Gaussian mixture model (GMM). In previous work, edema regions were segmented by selecting voxels with density values within the edema density distribution. This work improves upon the prior work by generating an adipose tissue mask without edema through a conditional generative adversarial network. The density distribution of the generated mask was imported into a Chan-Vese level set framework. Edema and subcutaneous adipose tissue are separated by iteratively updating their respective density distributions.

RESULTS

Validation results on 25 patients with edema showed that the segmentation accuracy significantly improved. Compared to GMM, the average Dice Similarity Coefficient increased from 56.0 to 61.7% ([Formula: see text]) and the relative volume difference decreased from 36.5 to 30.2% ([Formula: see text]).

CONCLUSION

The generated adipose tissue density prior improved edema segmentation accuracy. Accurate edema volume measurement may prove clinically useful.

A Dual-Branch Network with Mixed and Self-Supervision for Medical Image Segmentation: An Application to Segment Edematous Adipose Tissue

In clinical applications, one often encounters reduced segmentation accuracy when processing out-of-distribution (OOD) patient data. Segmentation models could be leveraged by utilizing either transfer learning or semi-supervised learning on a limited number of strong labels from manual annotation. However, over-fitting could potentially arise due to the small data size. This work develops a dual-branch network to improve segmentation on OOD data by also applying a large number of weak labels from inaccurate results generated by existing segmentation models. The dual-branch network consists of a shared encoder and two decoders to process strong and weak labels, respectively. Mixed supervision from both labels not only transfers the guidance from the strong decoder to the weak one, but also stabilizes the strong decoder. Additionally, weak labels are iteratively replaced with the segmentation masks from the strong decoder by self-supervision. We illustrate the proposed method on the adipose tissue segmentation of 40 patients with edema. Image data from edematous patients are OOD for existing segmentation methods, which often induces under-segmentation. Overall, the dual-branch segmentation network yielded higher accuracy than two baseline methods; the intersection over union (IoU) improved from 60.1% to 71.2% (p < 0.05). These findings demonstrate the potential of the dual-branch segmentation network with mixed- and self-supervision to process the OOD data in clinical applications.