Subcutaneous edema segmentation on abdominal CT using multi-class labels and iterative annotation

Abstract

Purpose: Anasarca is a condition that results from organ dysfunctions, such as heart, kidney, or liver failure, characterized by the presence of edema throughout the body. The quantification of accumulated edema may have potential clinical benefits. This work focuses on accurately estimating the amount of edema non-invasively using abdominal CT scans, with minimal false positives. However, edema segmentation is challenging due to the complex appearance of edema and the lack of manually annotated volumes.

Methods: We propose a weakly supervised approach for edema segmentation using initial edema labels from the current state-of-the-art method for edema segmentation (Intensity Prior), along with labels of surrounding tissues as anatomical priors. A multi-class 3D nnU-Net was employed as the segmentation network, and training was performed using an iterative annotation workflow.

Results: We evaluated segmentation accuracy on a test set of 25 patients with edema. The average Dice Similarity Coefficient of the proposed method was similar to Intensity Prior (61.5% vs. 61.7%; $p=0.83$ ). However, the proposed method reduced the average False Positive Rate significantly, from 1.8% to 1.1% ( $p<0.001$ ). Edema volumes computed using automated segmentation had a strong correlation with manual annotation ( $R^2=0.87$ ).

Conclusion: Weakly supervised learning using 3D multi-class labels and iterative annotation is an efficient way to perform high-quality edema segmentation with minimal false positives. Automated edema segmentation can produce edema volume estimates that are highly correlated with manual annotation. The proposed approach is promising for clinical applications to monitor anasarca using estimated edema volumes.

Keywords: Anasarca; Edema segmentation; Iterative annotation; Weakly supervised learning; nnU-Net.

Fig. 1

CT scans of male (A) and female patients (B), with edema present throughout the subcutaneous adipose tissue (yellow arrows). While edema has visible intensity differences with surrounding adipose tissue, it has complex shapes, unclear boundaries, and similar intensity characteristics with other tissues, e.g., glandular breast tissue in women (magenta arrows)

Fig. 2

Schematic of the proposed weakly supervised method for edema segmentation: A Initial multi-class labels on an abdominal CT scan of a female patient for training the nnU-Net and B the iterative annotation workflow

Fig. 3

Comparison of labels from Intensity Prior, a single-class nnU-Net and a multi-class nnU-Net. The presence of labels other than edema (yellow) in the multi-class model, such as for glandular breast tissue (cyan), mitigates false positives that otherwise appear with the Intensity Prior method and the single-class nnU-Net model

Fig. 4

Visual example of the improvement in edema segmentation at different stages of iterative annotation (Fig. 2B). False positives in the breast glandular tissue are progressively removed by re-training the nnU-Net with additional manually refined pseudo-labels

Fig. 5

Examples of edema segmentation produced by GMM, Intensity Prior method, and the proposed method on four anasarca patients. Each column (A–D) corresponds to a unique patient. The Intensity Prior method and the proposed multi-class nnU-Net method produce accurate segmentations generally. However, the Intensity Prior method generated visible regions of incorrect segmentation that are mitigated by the proposed method

Fig. 6

Comparison of a Dice Similarity Coefficient (DSC) and b False Positive Rate (FPR) metrics for edema segmentation. While the DSC metric could not capture differences in segmentation quality, the FPR metric distinguished the proposed method as the one to produce the lowest amount of false positive segmentations

Fig. 7

$R^2$ correlation plots of the automated edema segmentation methods compared against manual annotations. All three methods showed a strong correlation, with the Intensity Prior and proposed methods exhibiting a higher correlation ($R^2=0.87$) compared to GMM ($R^2=0.74$)