Fully Automated and Standardized Segmentation of Adipose Tissue Compartments by Deep Learning in Three-dimensional Whole-body MRI of Epidemiological Cohort Studies

Thomas Küstner, Tobias Hepp, Marc Fischer, Martin Schwartz, Andreas Fritsche, Hans-Ulrich Häring, Konstantin Nikolaou, Fabian Bamberg, Bin Yang, Fritz Schick, Sergios Gatidis, Jürgen MachannarXiv 2020 · arXiv

TL;DR

This paper introduces DCNet, a novel 3D deep learning architecture for automatically segmenting adipose tissue compartments from whole-body MRI. This method achieves high accuracy and speed, enabling rapid, objective, and standardized assessment of body composition for large-scale epidemiological studies, which is crucial for understanding metabolic diseases.

What Problem Does This Paper Address?

Accurately quantifying and localizing different types of adipose tissue (subcutaneous and visceral) from whole-body MRI is vital for understanding metabolic conditions like insulin resistance and type 2 diabetes. However, manual segmentation is time-consuming and prone to inter-observer variability. Existing automated methods often struggle with the variability in imaging sequences, scanners, and acquisition sites found in large epidemiological cohorts. This paper addresses the need for a robust, generalizable, and automated solution for whole-body adipose tissue segmentation that can handle diverse datasets.

Methods

The core of this work is the development of a 3D convolutional neural network (CNN) called DCNet. DCNet builds upon established architectures like UNet and VNet, incorporating several key innovations:

  • Merge-And-Run (MRGE) blocks: These blocks combine dense connections with attention-based multi-resolution focusing, allowing the network to effectively capture features at different scales.
  • Dense Convolutional Nodes: Inspired by DenseNet, these nodes facilitate deep supervision and efficient feature reuse.
  • Transition Layers: These layers handle downsampling (encoder) and upsampling (decoder) of feature maps using max-pooling and transposed convolutions.
  • Positional Encoding: To address potential loss of spatial information due to input patching, the absolute 3D position of each input cube is concatenated to the feature map in the deepest layer.

The network was trained using semi-supervised learning, starting with an automated fuzzy c-means clustering and snake algorithm to generate initial segmentations, which were then manually refined by experts.

Datasets

The study utilized data from two major epidemiological cohorts:

  • Tuebingen Family Study (TUEF) and German Center for Diabetes Research (DZD) database: This combined dataset provided 1000 cases, acquired on 1.5T and 3T scanners using T1-weighted fast spin echo (FSE) sequences.
  • German National Cohort (NAKO) database: This dataset contributed 300 cases, acquired on 3T scanners using multi-echo chemical-shift (Dixon) sequences.

These datasets were intentionally diverse in terms of imaging sequences, contrasts, scanners, field strengths, and patient positioning to test the robustness and generalizability of the proposed method.

Key Insights and Ideas

  • Robustness across diverse data: The DCNet demonstrated remarkable robustness, performing well across different imaging sequences (FSE vs. Dixon), scanners, field strengths (1.5T vs. 3T), and patient positioning (prone vs. supine). This is a significant step towards real-world clinical applicability.
  • Transfer learning is key for generalization: While intra-database training yielded the best results, transfer learning (training on one database and fine-tuning on another) and multi-database learning were crucial for overcoming performance drops when dealing with significantly different datasets (e.g., TUEF/DZD to NAKO).
  • Attention mechanisms improve segmentation: The integration of attention-based multi-resolution focusing within the MRGE blocks appears to be a critical factor in DCNet's superior performance compared to a standard 3D UNet, particularly in handling complex anatomical structures.
  • Speed is a game-changer: The ability to segment whole-body MRI in just 5-7 seconds post-loading is a major advantage, enabling seamless integration into clinical workflows and rapid phenotypic profiling.

Key Results

  • High Performance Metrics: On average, DCNet achieved excellent segmentation performance across all tissues and databases:
  • Sensitivity: 96.6%
    • Specificity: 95.1%
    • Precision: 92.1%
    • Accuracy: 98.4%
    • Dice Overlap: 0.94
  • Outperforms 3D UNet: DCNet significantly outperformed a standard 3D UNet across most metrics, with notable improvements in sensitivity (30.2%), specificity (40.1%), precision (151.6%), and Dice score (46.9%).
  • Fast Prediction: Once trained, the DCNet can predict segmentations in just 5-7 seconds, a substantial improvement over manual methods.
  • Qualitative Results: Visualizations show accurate segmentation of subcutaneous adipose tissue (SAT), visceral adipose tissue (VAT), and lean tissue (LT), even in challenging areas affected by magnetic field inhomogeneities. The generated adipose tissue profiles provide a clear head-feet visualization of fat distribution.

Limitations and Future Work

  • Specific to study design: The authors acknowledge that the results are specific to the imaging setup and MR sequence design used in their study. Generalizability to other segmentation tasks will require further investigation.
  • Bone marrow misclassification: The current network is not trained to distinguish bone marrow, leading to misclassifications in that region. However, for the primary task of adipose tissue profiling, this misclassification is minimal.
  • Future Work: The authors plan to extend the labeling and classification to include a bone marrow class. Further investigation into generalizability to other segmentation tasks and methods is also planned.

Summary updated

BibTeX
@misc{kustner2020fully,
  author = {Thomas Küstner and Tobias Hepp and Marc Fischer and Martin Schwartz and Andreas Fritsche and Hans-Ulrich Häring and Konstantin Nikolaou and Fabian Bamberg and Bin Yang and Fritz Schick and Sergios Gatidis and Jürgen Machann},
  title = {{Fully Automated and Standardized Segmentation of Adipose Tissue Compartments by Deep Learning in Three-dimensional Whole-body MRI of Epidemiological Cohort Studies}},
  year = {2020},
  eprint = {2008.02251},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2008.02251},
}
Figure