Deep Learning Diagnosis Method Of Muscle Atrophy Based On Ultrasound Multimodal Video

As the most common clinical disorder of skeletal muscle,muscle atrophy is mostly caused by muscle disease or nervous system dysfunction and typically occurs as a secondary consequence of disuse,acute and chronic illness and aging.Effectively and timely diagnosis of muscle atrophy is of great value for disease diagnosis and prognostic evaluation.As the preferred imaging method for diagnosing muscle atrophy,Ultrasonography(US)has the advantage of no radiation,low price and real time.This research focused on the unique mechanical properties of skeletal muscle,and developed deep learning method based on multimodal data including brightness mode(B-mode)and shear-wave elastography(SWE)for early diagnosis of muscle atrophy.the main contents are as follows:Muscle atrophy causes changes in the mechanical characterization of skeletal muscle.For the shortcomings that the mechanical characterization of muscle is difficult to be characterized by a single frame of SWE,we proposed to collect SWE videos of healthy subjects and patients with muscle atrophy when the gastrocnemius muscle is passively stretched.We conduct the preprocessed image sequence by means of feature learning.Then the extracted low-level and high-level features were fused,and output the probability distribution of muscle atrophy through the softmax function.The proposed method achieves an accuracy of 90.8%±0.034 for the recognition of muscle atrophy on the SWE videos,and the sensitivity and specificity of 90.3%±0.027 and91.4%±0.031,respectively.It shows great clinical promising application in terms of diagnosis of muscle disease.The B-mode US can image the anatomical structure of the muscle,obtain the structural parameters such as the muscle volume,thickness and cross-sectional area.Real-time SWE has been developed as a non-invasive and quantitative tool to measure elastic modulus of soft tissue.In Chapter 4,the automatic classification method for muscle atrophy,which is based on convolutional neural networks(CNN)under the Bmode,SWE,and dual-mode.The VGG neural network algorithm model based on deep learning was constructed and made appropriate improvements.We removed the fifth group of convolutions and fully connected layers,simplified the network by reducing the amount of network parameters and added a batch normalization algorithm to prevent overfitting during training.The results show that the model based on the dual-mode is significantly better than the single ultrasound mode in the in the diagnosis of muscle atrophy.We achieved the accuracy and sensitivity of 95.56% and 95.57%,respectively.The anatomical structure characteristics and elastic modulus of muscles through fusion are extracted by the algorithm,which further improves the accuracy of muscle atrophy,and realizes the effective discrimination of muscle atrophy.