Neural Architecture Search For Medical Image Aided Diagnosis

With the development of deep learning,deep neural network models have been widely utilized in intelligent Computer-Aided diagnosis(CAD)systems,where the neural architecture design heavily relies on the professional knowledge of artificial intelligence researchers.NAS technology aims to automatically design the high-performance neural networks,reduce the reliance on artificial intelligence experts in the construction of intelligent CAD systems,and improve the development efficiency and diagnosis performance of the systems.However,the current NAS methods are mainly designed for natural images,while there are significant differences in imaging equipment,collection difficulty,data distribution,and visual characteristics in the field of medical images.Medical image datasets usually have generally small but diverse data scales,imbalanced categories,and various imaging modalities.Applying traditional NAS technology directly to the field of medical imaging faces many challenges,and it could not achieve expected search performance.These challenges can be summarized in the following aspects:1.Generally small data scale——inaccurate evaluation modelThe annotation of medical images is highly dependent on medical expertise and requires a lot of effort and time from doctors.The high annotation cost leads to a generally small scale of medical image datasets.The performance impact of small data scale on the NAS methods mainly includes two aspects:1)The performance of neural architecture evaluation is crucial to the overall performance of NAS methods,and the evaluation model relies on training with a large amount of data.Too small data scale may weaken the prediction performance of the evaluation model.2)Most of the conventional NAS methods evaluate the performance of neural architectures based on the performance of verification sets.There are significant differences in the data distribution between training set and validation set based on the partitioning over small-scale data,which can indirectly lead to inaccurate architecture performance evaluation.In summary,small-scale medical datasets may lead to inaccurate evaluation model in NAS methods,which may affect the overall search effectiveness.2.Imbalanced categories——inaccurate evaluation metricMost medical image datasets are directly derived from real clinical scenarios,characterized by a long-tailed distribution.This can easily lead to the prediction of the model tending towards a poor recognition of tail category samples.In medical images,tail category samples often represent serious diseases,and accurate recognition of the tail categories is more crucial.Traditional NAS methods typically use metrics such as accuracy and cross entropy to evaluate neural architectures.These metrics can not accurately measure the performance of neural architectures.The NAS methods based on the above metrics may lead to a low disease detection rate in the CAD system,which cannot be applied to practical clinical scenarios.Therefore,the imbalanced nature of medical data categories can also lead to inaccurate architecture evaluation of traditional NAS methods,which can affect the overall search results.3.Diverse data scales between different datasets——requiring to dynamically search the model capacity based on the data scaleThe difficulty of collecting and labeling medical images is closely related to the disease type.The patient population size varies among different diseases,leading to significant differences in the natural distributions and data scales of different datasets,as well as the complexity of data samples.According to statistical machine learning theory,model capacity needs to be dynamically adjusted based on the data scale to achieve optimal performance on different datasets,otherwise it can easily lead to models falling into over-fitting or under-fitting.However,traditional NAS methods usually search in the search spaces with fixed network depths,without consideration of the network depths.These methods are not able to search for specific network depths based on data scale and adjust the model capacity dynamically.4.Different imaging methods and visual characteristics——requiring to adaptively search data augmentation strategies based on data typesMedical images involve different tissues,organs,and diseases,and they are generated through different imaging modalities.Their visual texture characteristics,lesion size,and distribution location vary greatly,making them very sensitive to data augmentation methods.Most traditional NAS methods only focus on the search of neural architecture,rarely considering the search of data augmentation strategies.A few relevant studies have explored the search of data augmentation strategies,but few of them consider the impact of data scale on the search of augmentation strategy.In addition,traditional methods usually utilize an independent way to search for data augmentation strategies and neural architectures,which is quite inefficient.In view of the common characteristics of the medical image field such as small data scale,imbalanced categories,diverse data scales,and various imaging modalities,this paper focuses on the issues such as the inaccurate evaluation model and evaluation metric when applying the traditional NAS technology to medical images,as well as the requirements for the search of neural network depth and data augmentation strategy,and conducts research on neural architecture search for medical image aided diagnosis.This research aims at improving the NAS evaluation accuracy,expanding the current search space,and improving the robustness,generalization,and practicality of NAS methods on medical image datasets.The specific research content and innovation are as follows:1.Improving the training method of arcihtecture evaluation modelTo address the problem of inaccurate evaluation model and evaluation metric caused by small scale and unbalanced categories of medical image datasets,and improve the architecture evaluation accuracy,this paper proposes an evaluation model training method for medical image datasets.Specifically,to address the issue of inaccurate model evaluation,this paper proposes a hierarchical and dynamical supernet pruning method during the search process,which continuously reduces the model capacity of the supernet and the generalization error of the model between the training set and the verification set,and then improves the accuracy of model evaluation.To address the issue of inaccurate evaluation metric,this paper proposes a search strategy based on multi-objective evolutionary algorithms,using the area under the receiver operating characteristic curve,which is more robust to category imbalance,as the main search objective to improve the robustness of evaluation metric against category imbalance.Finally,this paper utilize a multi-objective evolutionary algorithm to synthesize multiple optimization objectives.The network architecture searched on multiple medical datasets exhibits excellent performance,surpassing the artificially designed neural architecture and state-of-the-art NAS methods.2.Designing more effective architecture evaluation metricsAlthough the proposed pruning strategy-based evaluation model training method can improve the architecture evaluation and final search performance on medical image data,it still requires to train a supernet for architecture evaluation,which leads to an expensive search cost.At the same time,the evaluation method is still built based on the performance of validation set,inevitably suffering from the interference of small data scale and category imbalance of medical image dataset.In order to further improve search performance,this paper proposes a design method for architecture evaluation metric for medical image data,which utilizes randomly initialized models and various types of zero-cost metrics to evaluate neural architectures.Firstly,this paper proposes a general design principle and calculation paradigm for zero-cost metrics.This design principle can not only summarize existing advanced zero-cost metrics,but also can be used for mining more high-performance zero-cost metrics,greatly reducing the design difficulty.Then,this paper analyzes the computational time cost of different zero-cost metrics,and proposes an efficient computational method for them.This method can simultaneously calculate multiple zero-cost metrics in one forward and forward pass,significantly improving the evaluation efficiency.Finally,based on the newly designed zero-cost metrics and the proposed efficient computational method,this paper proposes an effective multi-objective evolutionary zero-shot neural network architecture search method,which has been evaluated on multiple medical datasets.The proposed can achieve competitive search performance and 20 times faster than state-of-the-art methods.3.Expanding the search over network depthTo address the issue of diverse data scale between different medical datasets,and search specific network depth to dynamically adjust model capacity,this paper proposes a neural architecture depth search method for medical image data.When searching for network depth,the current NAS methods usually adopt Bayesian neural networks to learn the uncertainty of network depth,and utilize variational inference to approximate the posterior distribution of network depth.Most of these methods are built based on the mean-field hypothesis.This paper firstly demonstrates that mean-field variational inference can lead to the inaccurate model evaluation,and then proposes a neural architecture depth search method based on structured variational inference,which can improve the approximate fidelity of the posterior distribution by modeling the neural weights with a condition on neural depth.The effectiveness of the proposed method has been evaluated on multiple medical datasets,demonstrating that the proposed method can search a reasonable neural architecture depth based on different data scales,further improving the practicality of NAS technology in intelligent CAD systems.In addition,the experiments on multiple natural datasets with simulated different data scales have demonstrated the generalization performance of the proposed method.4.Expanding the search over data augmentation strategyTo address the issue of medical image dataset such as diverse imaging modalities and the higher sensitivity on data augmentation strategies,this paper proposes a data augmentation search method for medical image data,which searches for specific data augmentation strategies according to the data type.Firstly,this paper demonstrates that search strategy based on density matching highly relies on a large data scale,while the small medical data scale can reduce the evaluation accuracy on augmentation strategies.In order to solve the problem,this paper proposes an automated data augmentation(auto-augmentation)method based on augmented density matching,which utilizes a random sampling strategy from a prior distribution of augmentation strategies to augment the training data and improve the evaluation accuracy of each augmentation strategy.Then,the current NAS methodology and auto-augmentation methodology are independent of each other,and the independent search way is inefficient.This paper proposes a unified search framework for data augmentation and neural architecture,which treats data augmentation operation and neural operation homogenously.Finally,the proposed method has been evaluated on multiple medical datasets to demonstrate its effectiveness.Based on the proposed search method,this paper has developed an intelligent skin CAD system for clinical diagnosis and treatment of acne.Through man-machine comparative experiments on clinical real dataset,the effectiveness of the constructed system has been evaluated.In summary,this paper focuses on the problems and challenges faced by neural architecture search in medical image aided diagnosis,aims at improving the prediction performance of evaluation model and the accuracy of evaluation metric,and expanding the search for neural architecture depth and data augmentation strategy.The method proposed of this paper can improve the search performance of NAS technology on small-scale and category-imbalanced medical data,dynamically search optimal model capacity based on the data scale and dataset complexity,adaptively search specific data augmentation strategies for different image types,and provide technical support for introducing NAS technology into the medical image domain.It is of great significance to further enhance the practicality of intelligent medical image aided diagnosis systems.