Personalized medicine, while prominent in oncology, is increasingly relevant for metabolic diseases like obesity and type 2 diabetes due to their complex genetic backgrounds and varying responses to treatments. Adipose tissue, specifically brown adipose tissue (BAT), plays a crucial role in energy expenditure. BAT activity is heterogeneous across the population, potentially influenced by genetics, and linked to metabolic health and cancer cachexia. Activating or suppressing BAT offers potential personalized treatment strategies for various metabolic diseases. However, current non-invasive quantification of BAT activity relies on [¹⁸F]-FDG PET scans, which are expensive and involve radiation exposure. This limits large-scale population studies. Predicting [¹⁸F]-FDG uptake from unenhanced computed tomography (CT) scans, which is less costly and less invasive, could significantly advance BAT research. Prior studies have shown a correlation between BAT's Hounsfield Unit (HU) in CT scans and its [¹⁸F]-FDG uptake in PET scans, motivating the development of computational methods for this prediction. This study explores the potential of Convolutional Neural Networks (CNNs) to predict BAT activity from CT scans, leveraging their ability to learn complex feature relationships without explicit feature engineering. The use of CNNs would facilitate larger cohort studies, enabling more effective patient stratification and personalized treatments.

Several studies have explored the relationship between brown adipose tissue (BAT) activity and metabolic health. Research has highlighted the heterogeneity of BAT presence and activity across individuals, suggesting a possible genetic component. Some studies link active BAT to a lower prevalence of cardio-metabolic diseases, while others indicate a role for BAT dysfunction in obesity. Furthermore, BAT deregulation is implicated in cancer cachexia. Personalized strategies targeting BAT activity, such as using bile acids or beta-agonists to activate BAT or propranolol to suppress it, are emerging. However, limitations exist in the size and genetic diversity of cohorts used in previous research. The current gold standard for assessing BAT activity, [¹⁸F]-FDG PET, is expensive and exposes patients to radiation. This has prompted research into alternative non-invasive methods, including the use of CT scans. Previous studies provided evidence of higher tissue density in BAT compared to white adipose tissue (WAT), resulting in higher CT intensity (HU). A study by Baba et al. showed a correlation between average HU and [¹⁸F]-FDG uptake (SUV) in supraclavicular BAT. This provides a basis for developing advanced computational methods to predict SUV from CT scans. CNNs, with their capacity for automatic feature learning, are promising for this task.

This study uses convolutional neural networks (CNNs) to predict the standardized uptake value (SUV) of brown adipose tissue (BAT) from unenhanced computed tomography (CT) scans. The focus is on the supraclavicular region, a major BAT depot in humans. The researchers trained a CNN using paired [¹⁸F]-FDG PET/CT datasets from four cohorts: two interventional cohorts with controlled cold stimulation to activate BAT and two retrospective clinical cohorts without cold stimulation. A total of 841 [¹⁸F]-FDG PET/CT scans from 718 subjects were used. The Attention U-Net architecture was employed. To assess the accuracy of the predictions, the researchers segmented active BAT regions from both predicted and actual PET scans using SUV thresholding. The Dice score was used to quantify the overlap between segmented regions. The performance of CNNs was compared to a HU thresholding-based method, a common technique for BAT segmentation from CT scans. The study also investigated whether CNN predictions could classify patients into BAT+ (high BAT activity) and BAT− (low BAT activity) groups. The area under the receiving operating characteristic curve (AUC) was used to evaluate the accuracy of this classification. Further analysis focused on the ability of CNNs to create stratified cohorts (cohorts with only BAT+ or BAT− subjects) using only CT scans. The number of subjects mistakenly included in each cohort was calculated to evaluate the performance of patient stratification. Qualitative results visualized the predicted and actual [¹⁸F]-FDG uptake of BAT. The researchers also explored the impact of dataset bias and the model's generalization ability across different cohorts. The datasets were preprocessed, including resampling to a consistent voxel size, cropping the region of interest around the supraclavicular region, and applying normalization techniques to CT and PET scans. For training, the researchers used 5 different random splits for training, validation, and testing. Data augmentation techniques were employed. The models were trained for 1000 epochs, using mean square error (MSE) as the loss function.

The study's key findings demonstrate that CNNs trained on cold-exposure cohorts significantly outperformed HU thresholding in segmenting active BAT regions. Intra-cohort experiments showed a substantial improvement in Dice scores: ~75% for the Basel cohort and ~23% for the Granada cohort. The improvement was statistically significant (p<<1%). However, performance was much lower in clinical cohorts (without cold stimulation). The CNN trained on the largest cold-exposure cohort (Granada) achieved a Dice score of 0.486 when tested on the Basel cohort and 0.538 on the Basel cohort, showing some degree of generalization. The Granada model demonstrated an ability to classify subjects into BAT+ and BAT− groups based on predicted BAT volume, achieving an AUC of ~0.8 compared to ~0.6 for HU thresholding. Furthermore, using CNN-predicted BAT volumes for patient stratification substantially reduced the number of mistakenly included subjects in stratified cohorts of BAT+ or BAT− compared to the HU thresholding method or random selection (~37% improvement). Qualitative results visually confirmed the accuracy of CNN predictions in many cases. The study found that training CNNs on retrospective clinical cohorts (without controlled cold exposure) was unstable due to ambiguities in mapping similar CT HU values to varying PET SUV values. Analysis revealed a negligible correlation between Dice score and BMI, suggesting that the model's predictive accuracy was not significantly confounded by variations in adiposity. The researchers also investigated the impact of using 2D versus 3D CNNs and determined that 2D CNNs were more effective with the data at hand. Segmentation accuracy improved for subjects with larger BAT volumes. The Granada model showed better generalization than the Basel model likely due to the dataset bias in the Basel cohort towards active BAT. Finally, The researchers successfully extended their approach to predicting BAT activity in other depots (cervical, paraspinal).

This study successfully demonstrates the potential of using CNNs to predict brown adipose tissue (BAT) activity from CT scans. The significant improvement in BAT segmentation accuracy and patient classification compared to traditional HU thresholding highlights the power of deep learning in this context. The ability to create stratified cohorts using only CT data represents a major advancement, significantly reducing the cost and radiation exposure associated with PET scans. This has important implications for large-scale studies aimed at understanding BAT's role in metabolic diseases. However, the limitations related to dataset bias and the instability of training on cohorts without cold stimulation need to be considered. Future research should focus on developing more robust methods for handling dataset imbalances and exploring alternative data acquisition protocols to minimize these limitations. The successful generalization to other BAT depots expands the applicability of this approach. The findings underscore the need for careful consideration of data acquisition protocols and potential biases when training and applying CNNs for medical image analysis.

This study demonstrates the feasibility of using convolutional neural networks (CNNs) to predict brown adipose tissue (BAT) activity from CT scans, significantly improving the accuracy of BAT segmentation and enabling more efficient patient stratification. This approach offers a promising path towards large-scale, cost-effective, and less invasive studies of BAT. Future work could explore more advanced network architectures (e.g., transformers), improved data acquisition protocols to mitigate dataset bias, and applications beyond BAT prediction.

The study's main limitations stem from the dataset biases present in the cold-exposure cohorts. This bias towards active BAT may have affected the generalizability of the models. The small size of some cohorts and variations in population characteristics between cohorts might have also impacted the generalization performance. Further research is needed to develop more robust models that can handle these biases and generalize well to diverse populations. The reliance on 2D CNNs might limit the accuracy of BAT segmentation compared to using 3D information; however, larger datasets are required for 3D CNN training to be effective.