Automated Brown Adipose Tissue Segmentation in CT Images of Lymphoma Patients

Project by Kasper Jørgensen

Introduction

Brown adipose tissue (BAT) plays a crucial role in energy expenditure and thermoregulation, making it a key focus in metabolic disease research. This project investigates the use of a nnU-Net model for automated BAT segmentation in CT images of lymphoma patients. By relying solely on anatomical CT data, the approach avoids potential biases linked to metabolic activity in PET-based analyses.

Project Background

BAT segmentation often combines imaging modalities such as PET/CT or PET/MRI. However, PET-based segmentation reflects metabolic activity, which may obscure the anatomical boundaries of BAT. To address this, the project focuses on segmenting BAT using only CT data, independent of metabolic signals.

Segmenting based solely on CT images is inherently challenging due to the minimal contrast between BAT and other adipose tissues. However, CT images still provide sufficient anatomical detail for physicians to identify BAT.

Manual segmentation is time-consuming and prone to variability, generating the need for automated solutions. Deep learning methods like the nnU-Net model offer a powerful alternative by directly mapping image data to segmentation masks, eliminating the need for handcrafted features. In this project, we employ a 3D nnU-Net model trained exclusively on CT data to segment BAT. Integration into the TotalSegmentator software allows effective BAT segmentation without PET data. The model has been trained and tested on 189 CT images from lymphoma patients.

Project Potential

This automated BAT segmentation model represents a significant step forward in BAT analysis, enabling reliable automated segmentation directly from CT scans without the need for PET data. While the current test set is limited, further validation with larger and more diverse datasets could improve its robustness and broaden its clinical applicability. By making BAT analysis more efficient, this approach can contribute to advancing metabolic research and optimizing clinical workflows.

Challenges Addressed by Automated CT-Based Segmentation
Figure 1: BAT Segmentation
Performance of the Automated BAT Segmentation Model
The BAT segmentations predicted by the model were compared against the ground truth annotations (the manually marked regions of BAT) in CT images.

To evaluate the model's accuracy, an agreement analysis was performed. This analysis helps visually identify where the model's predictions align with or diverge from the ground truth annotations.

Figure 1: BAT Segmentation
Figure 1 shows the BAT segmentation for four different test patients, comparing the original CT images, the manual ground truth annotations of BAT, the BAT segmentations predicted by the model, and the results of an agreement analysis.

True Positive (Green): Correctly predicted BAT by the model.

False Negative (Red):BAT missed by the model that should have been identified.

False Positive (Blue): Areas incorrectly identified as BAT, where there is no BAT.

The model's performance is not perfect, and inconsistencies are observed in areas where the ground truth annotations are unclear or less precisely defined.
Performance of the Model Against CT/PET
Figure 2: Results of Low-Activity Scanning
Figur 2 compares the predicted segmentations based solely on CT images with PET images for two patients with metabolically active BAT. The regions of increased PET activity align with the segmented BAT, suggesting that BAT can be accurately segmented using CT data alone, without the need for PET data.

Contact Information

Name: Kasper Jørgensen
Email: kasper.joergensen.02@regionh.dk
Location: Department of Clinical Physiology and Nuclear Medicine, Rigshospitalet, Denmark
Position: Student

Publications

Automated Supraclavicular Brown Adipose Tissue Segmentation in Computed Tomography Using nnU-Net: Integration with TotalSegmentator

GitHub