Weakly supervised learning for subcutaneous edema segmentation of abdominal CT using pseudo-labels and multi-stage nnU-Nets

Sayantan Bhadra ¹, Jianfei Liu ¹, Ronald M Summers ¹

⁰Imaging Biomarkers and Computer-Aided Diagnosis Laboratory, Radiology and Imaging Sciences, National Institutes of Health Clinical Center, Bethesda, USA.

Proceedings of Spie--the International Society for Optical Engineering +

|

October 7, 2024

View abstract on PubMed

Summary

This study introduces a weakly supervised deep learning method for accurate edema segmentation in CT scans. It improves upon existing techniques by using pseudo-labels, leading to more precise quantification of edema in clinical settings.

Area Of Science

Medical imaging analysis
Artificial intelligence in radiology
Quantitative medical imaging

Background

Accurate volumetric assessment of edema is crucial for monitoring diseases like kidney, liver, and heart failure.
Non-invasive, automatic edema segmentation from abdominal CT scans holds significant clinical value.
Current intensity-prior methods for edema segmentation suffer from false positives and under-segmentation.

Purpose Of The Study

To develop a weakly supervised learning method for 3D edema segmentation on abdominal CT scans.
To overcome limitations of manual annotation in supervised deep learning for edema segmentation.
To improve the accuracy and reliability of edema quantification for clinical applications.

Main Methods

Proposed a weakly supervised learning approach utilizing pseudo-labels derived from intensity priors for edema.
Incorporated pseudo-labels of surrounding tissues (muscle, subcutaneous and visceral adipose) for contextual information.
Employed a multi-stage nnU-Net architecture for refined 3D edema segmentation.

Main Results

The weakly supervised method demonstrated demonstrably lower segmentation errors compared to baseline methods.
Achieved more refined edema segmentations by leveraging pseudo-labels and contextual tissue information.
Validated the potential of the proposed approach for improved clinical edema quantification.

Conclusions

Weakly supervised learning with pseudo-labels offers a viable solution for 3D edema segmentation when manual annotations are scarce.
The integration of tissue pseudo-labels enhances segmentation accuracy by providing valuable context.
This method shows promise for advancing non-invasive edema assessment in various clinical conditions.

Keywords:

Anasarca edema segmentation nnU-Net pseudo-labels weakly supervised learning