Predicting standardized uptake value of brown adipose tissue from CT scans using convolutional neural networks

Area of Science:

• Medical Imaging
• Computational Biology
• Radiology

Background:

• Standard brown adipose tissue (BAT) identification uses [¹⁸F]-Fluorodeoxyglucose ([¹⁸F]-FDG) PET/CT, which is expensive and involves radiation exposure.
• This limits its practicality for large-scale population studies.
• Previous research indicates a correlation between CT Hounsfield Units (HU) of BAT and [¹⁸F]-FDG uptake, suggesting potential for computational prediction.

Purpose of the Study:

• To develop and validate a convolutional neural network (CNN) model for predicting [¹⁸F]-FDG uptake in BAT using only unenhanced CT scans.
• To assess the accuracy of CNN-based BAT segmentation compared to conventional CT thresholding.

Main Methods:

• Utilized the Attention U-Net architecture to train CNNs on datasets from four distinct cohorts.
• Focused predictions on restricted regions likely to contain BAT.
• Employed predicted [¹⁸F]-FDG uptake values for BAT segmentation.

Main Results:

• Achieved 23% to 40% higher accuracy in BAT segmentation compared to traditional CT thresholding methods.
• BAT volumes derived from CNN segmentation distinguished subjects with and without active BAT with an Area Under the Curve (AUC) of 0.8.
• CT thresholding yielded an AUC of 0.6 for the same distinction.

Conclusions:

• CNNs can effectively predict [¹⁸F]-FDG uptake in BAT from CT scans.
• This approach offers a more efficient and cost-effective alternative for large-scale imaging studies compared to PET/CT.
• CNN-based analysis enhances the ability to identify and quantify active BAT using only CT data.