An Semi-supervised Adversarial Learning Approach For Adipose Tissue Segmentation In Medical Images

Clinically,physicians often use visual CT images to grasp the patient’s adipose level.The visual judgment is often subjectively influenced by the doctor’s learning and experience,so this process carries the risk of misjudgment.Considering that medical images are characterized by high precision and high specialization,it is a hot issue in the field of medical image processing to obtain high precision segmentation models with limited labeled data by machines instead of human eyes.Some researchers have tried to train segmentation models semi-supervised using adversarial learning.However,such models are difficult to train and require high preprocessing of data,while their generalization performance is not satisfactory.Such segmentation models trained by small samples have not yet appeared in the field of adipose tissue segmentation.The research objectives of this paper are to find a adipose tissue segmentation method that can obtain high accuracy results based on small sample data.The thesis proposes a CT image localization method based on image classification,which can accurately find the CT slice that is near the third lumbar vertebra from the human CT image sequences obtained from CT scans.This slice can be used as an input to the adipose tissue segmentation method,which output an adipose tissue segmentation image.Experiments show that the method outperforms models such as FCN and VGG16 on the adipose tissue segmentation CT localization dataset.In addition,a semi-supervised adipose tissue segmentation method based on adversarial learning is proposed with small sample data as the entry point.The method processes CT medical images of human abdomen and is capable of segmenting subcutaneous adipose tissue and visceral adipose tissue simultaneously.In order to effectively improve the segmentation ability for the more complex visceral adipose tissue,a parallel task propagation mechanism is proposed in this thesis,which enables the subcutaneous adipose tissue segmentation task to guide the training of the visceral adipose tissue segmentation task.To further improve the accuracy of the model under semi-supervised training of small samples,this thesis proposes an untrustworthy region optimization mechanism that uses the redundant information obtained from adversarial learning to optimize the segmentation results of adversarial learning.Combining the above two mechanisms,the proposed fat segmentation method achieves an average Dice Cofficient of 95.208% in the abdominal fat dataset,which is better than other state-of-the-art adversarial learning methods.