0553

Automated Adipose Tissue Segmentation using 3D Attention-Based Competitive Dense Networks and Volumetric Multi-Contrast MRI

Sevgi Gokce Kafali^1,2, Shu-Fu Shih^1,2, Xinzhou Li¹, Shilpy Chowdhury³, Spencer Loong⁴, Samuel Barnes³, Zhaoping Li⁵, and Holden H. Wu^1,2
¹Radiological Sciences, University of California, Los Angeles, Los Angeles, CA, United States, ²Bioengineering, University of California, Los Angeles, Los Angeles, CA, United States, ³Radiology, Loma Linda University Medical Center, Loma Linda, CA, United States, ⁴Psychology, Loma Linda University School of Mental Health, Loma Linda, CA, United States, ⁵Medicine, University of California, Los Angeles, Los Angeles, CA, United States

Synopsis

Subcutaneous and visceral adipose tissue (SAT/VAT) are potential biomarkers to detect future risks of metabolic diseases. However, the current standard for analysis relies on manual annotations that require expert knowledge and are time-consuming. Previous neural networks for automatically segmenting adipose tissue had suboptimal performance for VAT. This work developed a new 3D attention-based competitive dense network to rapidly (84 ms/slice) and accurately segment SAT/VAT in adults with obesity by leveraging multi-contrast MRI inputs and considering the complex VAT features. The new network achieved high Dice scores (>0.96) and accurate volume measurements (difference<1.6%) for SAT/VAT with respect to manual annotations.

Introduction

The increased volume and fat content of subcutaneous and visceral adipose tissue (SAT, VAT) have strong associations with metabolic diseases ^{1, 2}. The reference standard to analyze SAT/VAT is manual annotations, which are time-consuming and may have inter-reader variability ^2-4. VAT is especially challenging due to its spatially varying complex nature. Previous deep learning methods for automated SAT/VAT segmentation used 2D slices or 3D patches as inputs and yielded suboptimal VAT segmentation performance ^5-7. These inputs were also limited to only T1-weighted in-phase and out-of-phase echoes (TE_IP, TE_OP) ⁵, or only fat (F) and water (W) images ⁶.
Recent work ⁸ proposed two 3-dimensional (3D) neural networks using full-volume multi-contrast inputs (TE_OP+W+F) to overcome the challenges associated with VAT: (1) 3D U-Net with weighted Dice loss (WDL) and (2) attention-based competitive dense 3D U-Net (ACD 3D U-Net) with a frequency-balancing boundary emphasizing Dice loss (FBDL) ⁹. Here, we further developed 3D U-Net and ACD 3D U-Net using a larger dataset of adults with obesity, performed ablation studies to assess the SAT/VAT segmentation performance, and strengthened the validity of the testing results.

Methods

Datasets: We included a subset of 629 subjects (mean body-mass-index: 33.00 kg/m²) from the IRB-approved Habitual Diet and Avocado Trial (HAT) ¹⁰. An axial T1-weighted 3D dual-echo Dixon sequence was acquired to measure SAT and VAT at 3T. TE_OP (1.23ms) and TE_IP(2.46ms) images were used to calculate F and W images ¹¹. Matrix size:192$$$\times$$$192, field-of-view (FOV): 400$$$\times$$$400 mm, 51 slices, slice thickness: 3.5 mm, TR: 5 ms, flip angle: 9$$$^{\circ}$$$. To be compatible with the 3D convolutional filter sizes, the number of slices was interpolated to 64. We stratified the dataset into training (70%), validation (10%), testing (20%) sets with similar characteristics.
Networks: We trained 3D U-Net ¹² with WDL ¹³, a common baseline network for SAT/VAT segmentation ⁶. Our proposed ACD 3D U-Net ⁸ consisted of competitive dense connections (Figure 1B) ¹⁴ and channel-spatial attention blocks (Figure 1C) ¹⁵. We used the FBDL ^{8, 9} to address the spatially varying, complex features of VAT. FBDL uses logistic weights ¹⁶ to address the class imbalance between SAT/VAT while emphasizing the boundaries ^{8, 9}. We used TE_OP, W, and F images as inputs to exploit the multi-contrast MRI information ⁸. The networks outputted SAT and VAT segmentations concurrently (Figure 1A).
Evaluations: Segmentation performance was evaluated using 3D Dice scores (Dice-VAT, Dice-SAT), false negatives (FN) [%], and false positives (FP) [%] per subject ¹⁷. Quantitative agreement in volume between manual annotations and network outputs was assessed using volume distance (VD) [%] per subject ¹⁷ (VD-SAT, VD-VAT).
Ablation study: We investigated three aspects. (1) Network structures: 3D U-Net with WDL vs. ACD 3D U-Net with WDL. (2) Loss functions: ACD 3D U-Net with WDL vs. FBDL. (3) Network inputs: ACD 3D U-Net with FBDL and W+F vs. TE_OP+TE_IN vs. TE_OP+W+F.

Results

Figures 2-3 show results for two representative subjects. ACD 3D U-Net yielded higher Dice-SAT and Dice-VAT for these subjects. FN and FP for ACD 3D U-Net were smaller than 3D U-Net (Fig. 4). Table 1 summarizes the ablation study results. All networks achieved accurate segmentation. The mean Dice-VAT improved from 0.947 (network 1) to 0.964 (network 3). Similarly, Dice-SAT increased for the proposed network (network 3). VD-SAT decreased from 1.2% (network 1) to 0.4% (network 3). VD-VAT was similar for networks 1 and 3. Compared to 3D U-Net (network 1), ACD 3D U-Net (networks 2-5) reduced the FN and FP.

Discussion

ACD 3D U-Net with FBDL outperformed 3D U-Net with WDL, achieving higher Dice scores, smaller FN, FP, and VD, and rapid inference times (84 ms/slice). Considering the smaller number of trainable parameters and superior performance, the proposed ACD 3D U-Net generalized the features better than 3D U-Net. More importantly, the proposed network markedly reduced gross errors, eliminating large FP regions in the SAT, arms, and vertebral bodies (Figs. 2-3). According to the ablation study (Table 1), the ACD blocks improved the performance the most. W and F images as inputs are critical for learning the informative features, as seen in the improvements for networks 2-4 vs. network 5.
Compared to previous works, our dataset is one of the largest (n=440 training, n=127 testing) and strengthened the validity of the results. Our work achieved higher Dice-VAT values (0.964) than the 2.5D FatSegNet with TE_OP+TE_IP (Dice-VAT: 0.85) ⁵ and 3D DCNet with W+F (Dice-VAT: 0.89) ⁶. These results demonstrated a promising advance in the state-of-the-art for VAT segmentation.
Our study had limitations. First, 3D networks require large memory due to full-FOV volumetric inputs. Future research can investigate the trade-offs between patch-based and full-FOV implementations. Second, the benefit of FBDL was not fully characterized as ACD may already capture the complex VAT details, especially with a large training dataset. Future work can analyze the advantages of FBDL for adapting the networks to datasets with varying characteristics and/or smaller size. Lastly, the differences in performance levels of previous works could be due to different subject characteristics and evaluation methodology.

Conclusion

ACD 3D U-Net segments SAT/VAT rapidly and accurately with state-of-the-art performance and could be an automated tool to analyze body composition in adults with obesity.

Acknowledgements

This study was supported in part by the Hass Avocado Board and the NIH (R01DK124417). This work was presented on behalf of the Habitual Diet and Avocado Trial (HAT) investigators.

References

1. Shuster, A., et al., The clinical importance of visceral adiposity: a critical review of methods for visceral adipose tissue analysis. Br J Radiol, 2012. 85(1009): p. 1-10.

2. Linge, J., et al., Body Composition Profiling in the UK Biobank Imaging Study. Obesity (Silver Spring), 2018. 26(11): p. 1785-1795.

3. Ly, K.V., et al., Free-breathing Magnetic Resonance Imaging Assessment of Body Composition in Healthy and Overweight Children: An Observational Study. J Pediatr Gastroenterol Nutr, 2019. 68(6): p. 782-787.

4. Bonekamp, S., et al., Quantitative comparison and evaluation of software packages for assessment of abdominal adipose tissue distribution by magnetic resonance imaging. International journal of obesity, 2008. 32(1): p. 100-111.

5. Estrada, S., et al., FatSegNet: A fully automated deep learning pipeline for adipose tissue segmentation on abdominal dixon MRI. Magnetic Resonance in Medicine, 2020. 83(4): p. 1471-1483.

6. Küstner, T., et al., Fully Automated and Standardized Segmentation of Adipose Tissue Compartments via Deep Learning in 3D Whole-Body MRI of Epidemiologic Cohort Studies. Radiology: Artificial Intelligence, 2020. 2(5): p. e200010.

7. Langner, T., et al., Fully convolutional networks for automated segmentation of abdominal adipose tissue depots in multicenter water–fat MRI. Magnetic resonance in medicine, 2019. 81(4): p. 2736-2745.

8. Kafali, S.G., et al., 3D Neural Networks for Visceral and Subcutaneous Adipose Tissue Segmentation using Volumetric Multi-Contrast MRI, in 43rd Annual Conference of IEEE Engineering in Medicine and Biology Society (EMBC). 2021.

9. Kafali, S.G., et al. A Densely Connected Neural Network with Frequency Balancing Loss for Adipose Tissue Segmentation in Children using Free-Breathing Abdominal MRI. in 29th Annual Meeting of International Society of Magnetic Resonance in Medicine. 2021.

10. Reboussin, D.M., et al., The design and rationale of a multi-center randomized clinical trial comparing one avocado per day to usual diet: The Habitual Diet and Avocado Trial (HAT). Contemporary Clinical Trials, 2021. 110: p. 106565.

11. Eggers, H. and P. Börnert, Chemical shift encoding‐based water–fat separation methods. Journal of Magnetic Resonance Imaging, 2014. 40(2): p. 251-268.

12. Çiçek, Ö., et al. 3D U-Net: Learning Dense Volumetric Segmentation from Sparse Annotation. in Medical Image Computing and Computer-Assisted Intervention – MICCAI 2016. 2016. Cham: Springer International Publishing.

13. Sudre, C.H., et al., Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations, in Deep learning in medical image analysis and multimodal learning for clinical decision support. 2017, Springer. p. 240-248.

14. Estrada, S., et al. Competition vs. concatenation in skip connections of fully convolutional networks. in International Workshop on Machine Learning in Medical Imaging. 2018. Springer.

15. Woo, S., et al. Cbam: Convolutional block attention module. in Proceedings of the European conference on computer vision (ECCV). 2018.

16. Roy, A.G., et al., QuickNAT: A fully convolutional network for quick and accurate segmentation of neuroanatomy. NeuroImage, 2019. 186: p. 713-727.

17. Taha, A.A. and A. Hanbury, Metrics for evaluating 3D medical image segmentation: analysis, selection, and tool. BMC medical imaging, 2015. 15(1): p. 29.

Figures

Figure 1. The network structure for ACD 3D U-Net. (A) We combined the TE_OP, water, and fat images as the multi-contrast input and outputted segmentation masks for SAT and VAT concurrently. (B) Each blue box represents a competitive densely connected block, which is more computationally efficient than a standard densely connected block. (C) At the bottom layer, we added convolutional attention blocks shown in dark gray color to consider the more informative spatial and channel features.

Figure 2. Input images (TE_OP, fat, and water), manual annotations, and output SAT (white) and VAT (gray) masks from 3D U-Net and ACD 3D U-Net in a representative subject. The Dice-VAT was higher for the ACD 3D U-Net (3D U-Net: 0.877, ACD 3D U-Net: 0.955) for this subject. The regions misclassified by 3D U-Net were indicated with red arrows. The corresponding regions (green arrows) were segmented correctly by ACD 3D U-Net. Manual annot.: manual annotations.

Figure 3. Input images (TE_OP, fat, and water), manual annotations, and output SAT (white) and VAT (gray) masks from 3D U-Net and ACD 3D U-Net in a second representative subject. For this subject, Dice-VAT for 3D U-Net and ACD 3D U-Net were 0.938 and 0.956, respectively. The regions misclassified by 3D U-Net were indicated with red arrows. The corresponding regions (green arrows) were segmented correctly by ACD 3D U-Net. Manual annot.: manual annotations.

Figure 4. False negatives (light blue) and false positives (orange) for two representative subjects using 3D U-Net and ACD 3D U-Net. For the example results shown here, ACD 3D U-Net achieved fewer false negatives and positives than 3D U-Net.

Table 1. Results of the ablation study in the testing dataset (n=127) in terms of 3D Dice scores, false negatives (FN), false positives (FP), and volume distance (VD). We compared network structure (networks 1 and 2), loss function (networks 2 and 3), and multi-contrast inputs (networks 3-5). The proposed ACD 3D U-Net with FBDL and (TE_OP+W+F) inputs is shown in gray background. The best results for each metric are indicated in green. All neural networks were trained on a NVIDIA Quadro RTX 8000 with 48 GB RAM. Results are reported as mean ± standard deviation (SD) across subjects. M: million.

Proc. Intl. Soc. Mag. Reson. Med. 30 (2022)

0553

DOI: https://doi.org/10.58530/2022/0553