Deep learning-based quantification of abdominal fat on magnetic resonance images

Automatic segmentation of large-scale CT image datasets for detailed body composition analysis

BACKGROUND

Body composition (BC) is an important factor in determining the risk of type 2-diabetes and cardiovascular disease. Computed tomography (CT) is a useful imaging technique for studying BC, however manual segmentation of CT images is time-consuming and subjective. The purpose of this study is to develop and evaluate fully automated segmentation techniques applicable to a 3-slice CT imaging protocol, consisting of single slices at the level of the liver, abdomen, and thigh, allowing detailed analysis of numerous tissues and organs.

METHODS

The study used more than 4000 CT subjects acquired from the large-scale SCAPIS and IGT cohort to train and evaluate four convolutional neural network based architectures: ResUNET, UNET++, Ghost-UNET, and the proposed Ghost-UNET++. The segmentation techniques were developed and evaluated for automated segmentation of the liver, spleen, skeletal muscle, bone marrow, cortical bone, and various adipose tissue depots, including visceral (VAT), intraperitoneal (IPAT), retroperitoneal (RPAT), subcutaneous (SAT), deep (DSAT), and superficial SAT (SSAT), as well as intermuscular adipose tissue (IMAT). The models were trained and validated for each target using tenfold cross-validation and test sets.

RESULTS

The Dice scores on cross validation in SCAPIS were: ResUNET 0.964 (0.909-0.996), UNET++ 0.981 (0.927-0.996), Ghost-UNET 0.961 (0.904-0.991), and Ghost-UNET++ 0.968 (0.910-0.994). All four models showed relatively strong results, however UNET++ had the best performance overall. Ghost-UNET++ performed competitively compared to UNET++ and showed a more computationally efficient approach.

CONCLUSION

Fully automated segmentation techniques can be successfully applied to a 3-slice CT imaging protocol to analyze multiple tissues and organs related to BC. The overall best performance was achieved by UNET++, against which Ghost-UNET++ showed competitive results based on a more computationally efficient approach. The use of fully automated segmentation methods can reduce analysis time and provide objective results in large-scale studies of BC.

Comparison of computed tomography and dual-energy X-ray absorptiometry in the evaluation of body composition in patients with obesity

Objective

a) To evaluate the accuracy of the pre-existing equations (based on cm2 provided by CT images), to estimate in kilograms (Kg) the body composition (BC) in patients with obesity (PwO), by comparison with Dual-energy X-ray absorptiometry (DXA). b) To evaluate the accuracy of a new approach (based on both cm2 and Hounsfield Unit parameters provided by CT images), using an automatic software and artificial intelligence to estimate the BC in PwO, by comparison with DXA.

Methods

Single-centre cross-sectional study including consecutive PwO, matched by gender with subjects with normal BMI. All the subjects underwent BC assessment by Dual-energy X-ray absorptiometry (DXA) and skeletal-CT at L3 vertebrae. CT images were processed using FocusedON-BC software. Three different models were tested. Model 1 and 2, based on the already existing equations, estimate the BC in Kg based on the tissue area (cm2) in the CT images. Model 3, developed in this study, includes as additional variables, the tissue percentage and its average Hounsfield unit.

Results

70 subjects (46 PwO and 24 with normal BMI) were recruited. Significant correlations for BC were obtained between the three models and DXA. Model 3 showed the strongest correlation with DXA (r= 0.926, CI95% [0.835-0.968], p<0.001) as well as the best agreement based on Bland - Altman plots.

Conclusion

This is the first study showing that the BC assessment based on skeletal CT images analyzed by automatic software coupled with artificial intelligence, is accurate in PwO, by comparison with DXA. Furthermore, we propose a new equation that estimates both the tissue quantity and quality, that showed higher accuracy compared with those currently used, both in PwO and subjects with normal BMI.

Phenotypic and Genetic Evidence for a More Prominent Role of Blood Glucose than Cholesterol in Atherosclerosis of Hyperlipidemic Mice

Hyperlipidemia and type 2 diabetes (T2D) are major risk factors for atherosclerosis. Apoe-deficient (Apoe−/−) mice on certain genetic backgrounds develop hyperlipidemia, atherosclerosis, and T2D when fed a Western diet. Here, we sought to dissect phenotypic and genetic relationships of blood lipids and glucose with atherosclerotic plaque formation when the vasculature is exposed to high levels of cholesterol and glucose. Male F2 mice were generated from LP/J and BALB/cJ Apoe−/− mice and fed a Western diet for 12 weeks. Three significant QTL Ath51, Ath52 and Ath53 on chromosomes (Chr) 3 and 15 were mapped for atherosclerotic lesions. Ath52 on proximal Chr15 overlapped with QTL for plasma glucose, non-HDL cholesterol, and triglyceride. Atherosclerotic lesion sizes showed significant correlations with fasting, non-fasting glucose, non-fasting triglyceride, and body weight but no correlation with HDL, non-HDL cholesterol, and fasting triglyceride levels. Ath52 for atherosclerosis was down-graded from significant to suggestive level after adjustment for fasting, non-fasting glucose, and non-fasting triglyceride but minimally affected by HDL, non-HDL cholesterol, and fasting triglyceride. Adjustment for body weight suppressed Ath52 but elevated Ath53 on distal Chr15. These results demonstrate phenotypic and genetic connections of blood glucose and triglyceride with atherosclerosis, and suggest a more prominent role for blood glucose than cholesterol in atherosclerotic plaque formation of hyperlipidemic mice.

Deep Learning-based Quantification of Abdominal Subcutaneous and Visceral Fat Volume on CT Images

RATIONALE AND OBJECTIVES

Develop a deep learning-based algorithm using the U-Net architecture to measure abdominal fat on computed tomography (CT) images.

MATERIALS AND METHODS

Sequential CT images spanning the abdominal region of seven subjects were manually segmented to calculate subcutaneous fat (SAT) and visceral fat (VAT). The resulting segmentation maps of SAT and VAT were augmented using a template-based data augmentation approach to create a large dataset for neural network training. Neural network performance was evaluated on both sequential CT slices from three subjects and randomly selected CT images from the upper, central, and lower abdominal regions of 100 subjects.

RESULTS

Both subcutaneous and abdominal cavity segmentation images created by the two methods were highly comparable with an overall Dice similarity coefficient of 0.94. Pearson's correlation coefficients between the subcutaneous and visceral fat volumes quantified using the two methods were 0.99 and 0.99 and the overall percent residual squared error were 5.5% and 8.5%. Manual segmentation of SAT and VAT on the 555 CT slices used for testing took approximately 46 hours while automated segmentation took approximately 1 minute.

CONCLUSION

Our data demonstrates that deep learning methods utilizing a template-based data augmentation strategy can be employed to accurately and rapidly quantify total abdominal SAT and VAT with a small number of training images.

Machine learning prediction of susceptibility to visceral fat associated diseases

Classifying subjects into risk categories is a common challenge in medical research. Machine Learning (ML) methods are widely used in the areas of risk prediction and classification. The primary objective of such algorithms is to use several features to predict dichotomous responses (e.g., healthy/at risk). Similar to statistical inference modelling, ML modelling is subject to the problem of class imbalance and is affected by the majority class, increasing the false-negative rate. In this study, we built and evaluated thirty-six ML models to classify approximately 4300 female and 4100 male participants from the UK Biobank into three categorical risk statuses based on discretised visceral adipose tissue (VAT) measurements from magnetic resonance imaging. We also examined the effect of sampling techniques on the models when dealing with class imbalance. The sampling techniques used had a significant impact on the classification and resulted in an improvement in risk status prediction by facilitating an increase in the information contained within each variable. Based on domain expert criteria the best three classification models for the female and male cohort visceral fat prediction were identified. The Area Under Receiver Operator Characteristic curve of the models tested (with external data) was 0.78 to 0.89 for females and 0.75 to 0.86 for males. These encouraging results will be used to guide further development of models to enable prediction of VAT value. This will be useful to identify individuals with excess VAT volume who are at risk of developing metabolic disease ensuring relevant lifestyle interventions can be appropriately targeted.

Classification of optical coherence tomography images using a capsule network

Background

Classification of optical coherence tomography (OCT) images can be achieved with high accuracy using classical convolution neural networks (CNN), a commonly used deep learning network for computer-aided diagnosis. Classical CNN has often been criticized for suppressing positional relations in a pooling layer. Therefore, because capsule networks can learn positional information from images, we attempted application of a capsule network to OCT images to overcome that shortcoming. This study is our attempt to improve classification accuracy by replacing CNN with a capsule network.

Methods

From an OCT dataset, we produced a training dataset of 83,484 images and a test dataset of 1000 images. For training, the dataset comprises 37,205 images with choroidal neovascularization (CNV), 11,348 with diabetic macular edema (DME), 8616 with drusen, and 26,315 normal images. The test dataset has 250 images from each category. The proposed model was constructed based on a capsule network for improving classification accuracy. It was trained using the training dataset. Subsequently, the test dataset was used to evaluate the trained model.

Results

Classification of OCT images using our method achieved accuracy of 99.6%, which is 3.2 percentage points higher than that of other methods described in the literature.

Conclusion

The proposed method achieved classification accuracy results equivalent to those reported for other methods for CNV, DME, drusen, and normal images.