| Medical ultrasound is a widely used imaging modality in clinical practice.Because of its unique advantages such as real-time,cheap and noninvasive,it has been developing rapidly.In recent years,with the development of deep learning in the field of computer vision,deep learning has also been widely bringing into ultrasound image analysis and achieved great success such as in lesions/nodules classification,organ segmentation and fetal standard plane detection.Deep learning based methods have achieved performance that surpasses traditional manual feature extraction.However,the deep neural networks used in traditional deep learning are mostly manually designed by experts which is a trial-error process that requires great experience and is time consuming and labor-intensive.Deep neural networks designed by hand often have large redundancy,such as parameters and model size.Deep neural networks design in the field of medical image analysis mostly draws on successful experiences in natural image preprocessing,such as Resnet and mask-RCNN,and uses transfer learning to apply to specific tasks,but there are similar problems.Auto machine learning aims to reduce labor costs in data-driven deep learning with auto data clean,auto model selection,auto hyper-parameter optimization and neural architecture search.As a part of auto machine learning,neural architecture search has received increasing attention.Neural architecture search is to select a set of parameters from the network architecture parameter space on the specified task and dataset,so that the obtained neural network has higher performance and fewer parameters.This thesis proposes a hybrid neural architecture search method that combines automatic neural architecture search with traditional neural architecture experience design,and incorporates transfer learning and gradient descent search algorithms.In the search process,the training dataset is divided into a training set and a validation set,and the network weight parameters and structure parameters are updated through bi-level optimization.In the search strategy,a progressive search method is used to gradually increase the network depth,reduce the search space redundancy,and improve the search efficiency.In this paper,experimental verification is performed on two basic tasks: classification of liver echinococcosis disease and ovary/follicle segmentation in ultrasound images.In the classification task,we first analyze the parameter amount and runtime resource occupation of each module of the baseline network,then replace the most redundant modules with the super network superimposed by the basic network Cell,and add the newly proposed Mixture convolution layer to the Cell operation set.In the verification phase,densely connection is used to form a Dense Cell to improve gradient conduction efficiency.In the segmentation task,we search the Decoder based on the lightweight network DeepLab-v3+ using the NAS method.By extending the ASPP multi-scale feature fusion module,the ASPP Cell is further enhanced to feature fusion.Comprehensive experimental results show that on classification tasks,the hybrid neural network architecture search makes the searched network reduce the parameters by 50% compared to the basic network,and the classification F1 score is increased by 1%.After integrating the traditional Dense connection idea,the performance is further improved.On the segmentation,this method reduces the network parameters by 50% compared to the lightweight baseline network.The new network has 12 M parameters and the dice coefficient increased by 0.5%.Neural network architecture search has a performance that surpasses the manually designed networks,which will be the trend and hotspot.This study tried to apply it to medical ultrasound image analysis and achieved state-of-the-art performance.This shows the advanced nature and great potential of the method. |
