Body Composition Profiling Related Medical Imaging Segmentation

BackgroundThe epidemic of obesity and other nutrition related chronic diseases has been one major public health challenge to most countries including China.Anthropometric measures like waist circumference and body mass index are indirect raw proxy of human body fat.Not a few studies reported inconsistent research results for various diseases based on these metrics,suggesting they are not right choices for assessing body fat.Additionally,they are unable to reflect the distributions of other body components like muscle.In recent years,body composition profiling,known as an objective and fine nutritional status assessment approach,has gained increasing popularity among nutritionists and physiologists.At the tissue-system level,body compositions comprise tissues including adipose tissue and muscle,organs,and systems,which represent distinct physiological functions.Different mass and distributions of the body compositions reflect differed status of health or diseaseComputed tomography(CT)and magnetic resonance imaging(MRI)are golden standards for detecting soft tissues and organs.The procedure of body composition profiling based on CT and MRI basically consist of three steps:image acquisition,body composition segmentation on the image,calculation of volume and mass of the body composition using the segmentation.Currently,the second step is highly depended on manual delineation by experts,which is time consuming and laborious and has been the bottleneck to widely apply CT and MRI to body composition analysisDeveloping efficient and accurate automatic segmentation tools will wipe out the above-mentioned technical problem for body composition profiling,facilitating its application in more settings.Therefore,we conducted two studies in sequence automatic segmentation of skeletal muscle and adipose tissues in MRI,and automatic segmentation of multiple organs in CT and MRI.MethodsTo segment the tissues,a convolutional network named MS-denseUNet was proposed based on U-Net.To minimize loss of details of the image,dense connection inspired by DenseNet was innovatively introduced to densely connect multi-scale feature maps.Two datasets were used,including 70 participants with MRI scanned in the second affiliated hospital of Zhejiang University during Dec 2009 and July 2010,and 36 participants with MRI scanned in the provincial hospital of Zhejiang during Oct 2012 and June 2013,respectively.The first dataset was used to train models,while the second was employed for external testing.Comparisons were performed with U-Net and CE-NetTo meet the practical needs to segment multiple organs simultaneously,on the ground of a traditional convolution network for segmentation,a universal network architecture called 3D U2-Net was built by replacing the standard convolution with an adapter module,which was self-designed based on separable convolution and contains both domain-specific parameters and domain-sharing parameters.A total of 6 datasets were utilized,5 of which,called base datasets,are focusing on liver,left atrium,pancreas,prostate and hippocampus,and were used to train and test the universal model The last dataset,called new dataset,contains spleen and was used to evaluate the adaptiveness of models.A universal model was acquired by training the universal network on five base datasets at once.For comparison,independent models were built by training the basic network on each base dataset and a shared model was built by training the basic network on all base datasets simultaneously.Dice coefficients(%)was used to evaluate the segmentation performance.ResultsMS-denseUNet achieved good performance for the studied tissues on test data,with Dice coefficients(%)of 86.19,88.01,72.50,55.93 for skeletal muscle,subcutaneous adipose tissue,visceral adipose tissue and intermuscular adipose tissue,respectively.MS-denseUNet performed consistently better for all tissues than U-Net.Compared to CE-Net,it was slightly worse for subcutaneous adipose tissue,but was much better for visceral adipose tissue and intermuscular adipose tissue.The universal model 3D U2-Net obtained comparable or even better performance compared to the independent models and the shared model with much less parameters.Specifically,the universal model performed much better than the independent model on Prostate,with Dice coefficients(%)of peripheral zone and transition zone being 68.50 and 89.21,respectively.Its performance on Pancreas(Dice coefficient(%)=62.08)was a little inferior to the independent model but it outperformed the shared model.3D U2-Net was the smallest,ratios of whose number of parameters to that of the independent models and shared model were 1%and 6%,respectively.What is more,3D U2-Net was better than the shared model on new dataset for spleen segmentation,showing superior generalizability.ConclusionWe successfully developed two models for fully automatic segmentation of body composition,MS-denseUNet was for whole body skeletal muscle and adipose tissues segmentation in MRI,and 3D U2-Net was a universal model for multi-organ segmentation regardless of CT or MRI.Both models generalized well to new datasets.In particular,the universal model for multi-organ segmentation contains much less parameters than the traditional methods,requiring less computation and storage memory,therefore it is a be more suitable to be deployed and applied in real-world scenarios.