Brown adipose tissue machine learning nnU-Net V2 network using TriDFusion (3DF)

Fig. 1. Visualization of segmentation exclusion zones. A) 3D volume rendering of the CT scan, along with machine learning organ segmentation, is used to compute different classification zones. The various organs are distinguished through the use of different colors. B) 3D MIP of the PET scan, highlighting brown fat in color and the exclusion zones in red

Fig. 2. Visualization of BAT classification. A) 3D volume rendering of the CT with machine learning organ segmentation used for brown fat zone classification. B) 3D MIP of the PET scan, with the computed Euclidean classification zones represented by different colors

Fig. 3. Semi-automatic workflow diagram. A) 3D MIP rendering of the PET AC of the patient. B) 3D volume rendering illustrating the CT AC of the patient. C) 3D volume rendering showcasing organ exclusion and zone exclusion. D) 3D volume rendering demonstrating zone classification. E) 3DF contour editing workflow. Using the 3DF contour review panel, the physician reviews each VOI by navigating to previous, deleting, or moving to the next contour. The classification associations for the VOI can also be changed as needed for each VOI. The refined contours are exported into a DICOM RT Structure set file and used to create the machine learning network

Fig. 4. PET/CT segmentation views. A) 3D MIP rendering of the PET AC fused with the segmentation of the different anatomical sites in color. B) 2D view of the fusion between a slice of the CT AC and PET images, displaying the segmentation of the different sites in that slice. C) Zoomed 2D view showing the segmentation of supraclavicular uptake. For the delineation, only uptake visible on both PET and CT fusion was used to define its boundary. D) 2D view of the CT image alone as in B, but without fusion with the PET image. E) Zoomed view of the CT image of the same delineation in C but without fusion with the PET image

Fig. 5. Per-reviewer rating distributions by anatomical site. Rating frequencies chart by reviewer across 118 cases per anatomical site. Ratings are color-coded from 1 to 5, where 5 indicates perfect performance, 4 great, 3 good, 2 fair, and 1 poor. The supraclavicular and abdominal regions exhibited the highest frequency of perfect ratings, followed by the axillary, paraspinal, cervical, and mediastinal regions

Fig. 6. Brown adipose tissue (BAT) activity (log10(SUVmean x volume)) versus weight, height, age, and BMI in females and males. Data from 3,883 segmented patients. Female values are plotted in pink and male values in blue, each with its own linear trendline with 95% confidence intervals. The violin plot depicts the distributions for males and females, with the median indicated by a yellow line and the 25th and 75th percentiles shown in black