Background

There is increasing appreciation of the association of obesity beyond co-morbidities, such as cancers, Type 2 diabetes, hypertension, and stroke to also impact upon the muscle to give rise to sarcopenic obesity. Phenotypic knowledge of obesity is crucial for profiling and management of obesity, as different fat—subcutaneous adipose tissue depots (SAT) and visceral adipose tissue depots (VAT) have various degrees of influence on metabolic syndrome and morbidities. Manual segmentation is time consuming and laborious. Study focuses on the development of a deep learning-based, complete data processing pipeline for MRI-based fat analysis, for large cohort studies which include (1) data augmentation and preprocessing (2) model zoo (3) visualization dashboard, and (4) correction tool, for automated quantification of fat compartments SAT and VAT.

Methods

Our sample comprised 190 healthy community-dwelling older adults from the Geri-LABS study with mean age of 67.85 ± 7.90 years, BMI 23.75 ± 3.65 kg/m², 132 (69.5%) female, and mainly Chinese ethnicity. 3D-modified Dixon T1-weighted gradient-echo MR images were acquired. Residual global aggregation-based 3D U-Net (RGA-U-Net) and standard 3D U-Net were trained to segment SAT, VAT, superficial and deep subcutaneous adipose tissue depots (SSAT and DSAT). Manual segmentation from 26 subjects was used as ground truth during training. Data augmentations, random bias, noise and ghosting were carried out to increase the number of training datasets to 130. Segmentation accuracy was evaluated using Dice and Hausdorff metrics.

Results

The accuracy of segmentation was SSAT:0.92, DSAT:0.88 and VAT:0.9. Average Hausdorff distance was less than 5 mm. Automated segmentation significantly correlated R² > 0.99 (p < 0.001) with ground truth for all 3-fat compartments. Predicted volumes were within ± 1.96SD from Bland–Altman analysis.

Conclusions

DL-based, comprehensive SSAT, DSAT, and VAT analysis tool showed high accuracy and reproducibility and provided a comprehensive fat compartment composition analysis and visualization in less than 10 s.

中文翻译：

---

CAFT：用于大型队列研究的基于深度学习的综合腹部脂肪分析工具

背景

人们越来越认识到肥胖与癌症、2 型糖尿病、高血压和中风等并发症之间的关联也会影响肌肉，从而导致少肌性肥胖。肥胖的表型知识对于分析和管理肥胖至关重要，因为不同的脂肪——皮下脂肪组织库 (SAT) 和内脏脂肪组织库 (VAT) 对代谢综合征和发病率有不同程度的影响。手动分割既费时又费力。研究的重点是开发基于深度学习的完整数据处理管道，用于基于 MRI 的脂肪分析，用于大型队列研究，包括 (1) 数据增强和预处理 (2) 模型动物园 (3) 可视化仪表板和 (4) ) 校正工具，用于自动量化脂肪隔室 SAT 和 VAT。

方法

我们的样本包括来自 Geri-LABS 研究的 190 名健康的社区居住老年人，平均年龄为 67.85 ± 7.90 岁，BMI 为 23.75 ± 3.65 kg/m ²，132 名（69.5%）女性，主要是华人。获得了 3D 修改的 Dixon T1 加权梯度回波 MR 图像。对基于剩余全局聚合的 3D U-Net (RGA-U-Net) 和标准 3D U-Net 进行了训练，以分割 SAT、VAT、浅表和深层皮下脂肪组织库（SSAT 和 DSAT）。在训练期间使用来自 26 个受试者的手动分割作为基本事实。进行数据增强、随机偏差、噪声和重影以将训练数据集的数量增加到 130 个。使用 Dice 和 Hausdorff 指标评估分割精度。

结果

分割精度为SSAT：0.92、DSAT：0.88和VAT：0.9。平均豪斯多夫距离小于 5 毫米。自动分割显着相关R ²  > 0.99 ( p  < 0.001) 与所有 3 脂肪隔间的基本事实。Bland-Altman 分析的预测体积在 ± 1.96SD 以内。

结论

基于 DL 的综合 SSAT、DSAT 和 VAT 分析工具显示出很高的准确性和可重复性，并在不到 10 秒的时间内提供了全面的脂肪室成分分析和可视化。