Related Experiment Video
Updated: Aug 28, 2025
06:03
Integrating Augmented Reality Tools in Breast Cancer Related Lymphedema Prognostication and Diagnosis
Published on: February 6, 2020
6.7K
Deep learning-based quantitative estimation of lymphedema-induced fibrosis using three-dimensional computed
Hyewon Son1, Suwon Lee1, Kwangsoo Kim2
1Major of Biomedical Engineering, Department of Electrical, Electronic and Computer Engineering, University of Ulsan, Nam-gu, Ulsan, 44610, Republic of Korea.
Scientific Reports
|September 13, 2022
Summary
Deep learning (DL) shows promise for recognizing lymphedema-induced fibrosis in CT scans. This new method may help detect fibrosis before it worsens, improving patient outcomes.
Area of Science:
- Medical Imaging
- Artificial Intelligence in Medicine
- Lymphedema Research
Background:
- Lymphedema involves progressive fibrosis driven by proinflammatory cytokines.
- Current methods for measuring lymphedema-induced fibrosis are limited, especially before significant deterioration.
- Computed Tomography (CT) can visualize fibrosis in both superficial and deep tissues.
Purpose of the Study:
- To verify the efficacy of deep learning (DL) for standardized measurement of lymphedema-induced fibrosis using CT images.
- To develop and validate novel indices for quantifying fibrosis based on DL segmentation.
Main Methods:
- A cross-sectional, observational cohort study analyzed 2138 CT images from 27 chronic unilateral lymphedema patients.
- A SegNet-based deep learning model was trained for semantic segmentation of CT images into five classes: air, skin, muscle/water, fat, and fibrosis.
- Four indices were formulated from the segmented images and compared with standardized circumference difference ratio (SCDR) and bioelectrical impedance (BEI).
Main Results:
- The DL model achieved a mean boundary F1 score of 0.868 and accuracy of 0.776 for fibrosis segmentation.
- A significant percentage of the formulated subindices (73.7%) correlated with BEI measurements (partial correlation coefficient: 0.420-0.875).
- A smaller proportion (13.2%) correlated with SCDR (0.406-0.460), with a specific mean subindex of Index 2 showing the highest correlation.
Conclusions:
- Deep learning demonstrates significant potential for recognizing and quantifying lymphedema-induced fibrosis from CT imaging.
- Subtraction-type formulas derived from DL analysis appear to be a promising method for fibrosis estimation.
- This approach could enable earlier detection and more accurate monitoring of lymphedema fibrosis.
