Introduction

Identification of distribution and type of adipose tissue (AT) in the human body plays an important role in the pathogenesis of metabolic diseases¹. The comparison of AT distribution in subgroups with different anthropometric parameters and metabolic conditions is of interest in order to identify and quantify high-risk subjects as well as to establish patient-specific treatments. In order to determine the risk of diabetes, the spatial distribution and quantity of subcutaneous adipose tissue (SCAT) and visceral adipose tissue (VAT) gives a good indicator². Especially in large cohort studies, manual tissue delineation (labeling) demands time- and cost-intensive efforts by clinical experts which is not feasible. Therefore, a reliable and automatic assessment is of great interest. Furthermore, the acquired MR images may suffer under magnetic field inhomogeneity artifacts which originate from field strength dependency due to bias-field non-uniformity resulting in varying image intensity. These artifacts are especially prominent in systems with field strengths >1.5T. Thus, correct tissue classification can be hampered and classical segmentation algorithms are misguided by this intensity variation. In order to overcome the limitation of previously described approaches for this task^3,4,5, i.e. intensity and boundary based methods, we propose the usage of a 3D convolutional neural network (CNN) for AT segmentation and determination of whole-body AT distribution. We utilize our previously proposed DCNet^6,7 and extend it by merge-and-run blocks to capture more multi-resolutional features simultaneously for accurate whole-body AT classification. Once trained the network can estimate the AT distribution markedly faster than previously described methods^3,4,5 saving time and costs.

Material and Methods

Anisotropic 2D transverse T1w FSE whole-body data sets¹ were acquired in a multi-center study on several different 1.5T (Siemens MAGNETOM Sonata Vision and Avanto) and 3T MR systems (Siemens MAGNETOM Prisma and Vida) covering the entire body in prone stretched position and allowing fast whole-body coverage. All data sets were semi-automatically labeled in four classes: background (BG), lean tissue (LT), SCAT and VAT. A classical boundary and intensity based approach³ was used for this and afterwards, all segmentation masks were inspected and corrected manually by skilled experts. The proposed neural network is based on our 3D DCNet approach^6,7 which consists of an encoding and decoding branch with convolution and deconvolution layers for feature map down-/upscaling in the respective branches as shown in Fig. 1. Each branch consists of four stages. Shortcut connections between branches and inside stages forward residual information. In contrast to the previous DCNet, each stage consists of merge-and-run mapping blocks⁸. Each merge-and-run mapping block consists of parallel branches with two 3D convolutions, rectified linear unit activation and batch normalization. This assembles several residual branches in parallel by adding the average of the input branches to the subsequent residual branch, and thus directly leading to a reduced network depth. The receptive field of the merge-and-run blocks was globally enhanced by adding global residual branches to the local residual paths by applying dilated convolutional kernels. This allows dealing with the single-channel input in a robust way and enables a wider multi-resolutional receptive field whilst also avoiding pooling layers. Input images were cropped from 256×178×98-120 to 32×32×32 overlapping blocks. The output into the four classes (background, lean tissue, SCAT, VAT) is generated by a finalizing layer with softmax activation function. Network weights were optimized by RMSprop with multi-class Focal Loss⁹ and multi-metric (true-positive, recall and Jaccard index), batch-size 48, 60 epochs. Out of 2000 subjects the neural network was trained on 300 randomly selected subjects (1.5T: 150, 3T: 150) with 80 validation and 306 test subjects (1.5T: 100, 3T: 206). Different training cases were conducted to evaluate the influence of intensity inhomogeneities on the classification result (Tab. 1).

Results and Discussion

Fig. 2 and 3 show exemplary results of the CNN segmentation masks color-coded overlaid to the MR image as well as the obtained recall per class and accuracies over all test subjects. Visual inspections of the segmentation output reveal good agreement to the manual labeled ground truth. Estimation is more robust to intensity inhomogeneities if 3T data is considered in training. Training of the CNN required 25h whereas only 5-7s are needed for label prediction after data set loading – compared to 15min in the classical boundary and intensity based approach. AT distribution over the slices shows good agreement with ground truth. Fig. 4 depicts gender-specific tissue distributions in head-feet direction for all male and female subjects.

Conclusion

Automatic, robust and fast AT segmentation together with standardized topography mapping is feasible in whole-body T1-weighted image data sets by convolutional neural networks.

Acknowledgements

No acknowledgement found.

References

1. Machann J, Thamer C, Schnoedt B, Haap M, Haring HU, Claussen CD, Stumvoll M, Fritsche A, Schick F. Standardized assessment of whole body adipose tissue topography by MRI. Journal of Magnetic Resonance Imaging: An Official Journal of the International Society for Magnetic Resonance in Medicine 2005;21(4):455-462.
2. Storz C, Heber SD, Rospleszcz S, Machann J, Sellner S, Nikolaou K, Lorbeer R, Gatidis S, Elser S, Peters A, Schlett CL, Bamberg F. The role of visceral and subcutaneous adipose tissue measurements and their ratio by magnetic resonance imaging in subjects with prediabetes, diabetes and healthy controls from a general population without cardiovascular disease. The British journal of radiology 2018;91(1089):20170808.
3. Würslin C, Machann J, Rempp H, Claussen C, Yang B, Schick F. Topography mapping of whole body adipose tissue using a fully automated and standardized procedure. Journal of Magnetic Resonance Imaging: An Official Journal of the International Society for Magnetic Resonance in Medicine 2010;31(2):430-439.
4. Karlsson A, Rosander J, Romu T, Tallberg J, Grönqvist A, Borga M, Dahlqvist Leinhard O. Automatic and quantitative assessment of regional muscle volume by multi‐atlas segmentation using whole‐body water–fat MRI. Journal of Magnetic Resonance Imaging 2015;41(6):1558-1569.
5. Addeman BT, Kutty S, Perkins TG, Soliman AS, Wiens CN, McCurdy CM, Beaton MD, Hegele RA, McKenzie CA. Validation of volumetric and single‐slice MRI adipose analysis using a novel fully automated segmentation method. Journal of Magnetic Resonance Imaging 2015;41(1):233-241.
6. Küstner T, Müller S, Fischer M, Weiβ J, Nikolaou K, Bamberg F, Yang B, Schick F, Gatidis S. Semantic Organ Segmentation in 3D Whole-Body MR Images. 2018. IEEE. p 3498-3502.
7. Küstner T, Fischer M, Müller S, Gutmann D, Nikolaou K, Bamberg F, Yang B, Schick F, Gatidis S. Automated segmentation of abdominal organs in T1-weighted MR images using a deep learning approach: application on a large epidemiological MR study. Proceedings of the International Society for Magnetic Resonance in Medicine (ISMRM). Paris, France2018.
8. Zhao L, Wang J, Li X, Tu Z, Zeng W. Deep convolutional neural networks with merge-and-run mappings. arXiv preprint arXiv:161107718 2016.
9. Lin T, Goyal P, Girshick R, He K, Dollár P. Focal Loss for Dense Object Detection. 2017 22-29 Oct. 2017. p 2999-3007.