Research On Cross-domain Adaptive Learning Methods For Medical Image Recognition

The emergence and evolution of deep learning has created new prospects for the advancement of medical image processing.The ideal scenario for deep learning technology is the existence of abundant labeled training instances,and it requires that these training instances have the same feature distribution as the test ones.However,in actual medical scenarios,due to factors such as different types of image acquisition equipment,different parameter settings,different models,and different scanning methods,there are significant differences in feature distribution between different medical images.As a result,deep learning methods are unsuitable for direct application in clinical practice.The proposal of cross-domain adaptive learning methods can effectively address the dependence of deep learning methods on feature distribution.However,due to the complexity and clinical application characteristics of cross-domain analysis of medical images,related work on cross-domain adaptive learning methods for medical images is still immature.In order to better overcome the feature distribution differences of medical images in clinical application scenarios,this paper conducts research on deep cross-domain adaptive learning methods for single-modal and multi-modal medical images.The specific research contents are as follows:1.Landmark detection of single-modal medical images based on the domain adaptation metric learning networkExisting methods for cross-domain landmark detection of single-modal medical images often overlook the impact of image feature distribution shifts on detection accuracy due to individual pose differences during image acquisition.To address these challenges,this paper proposes a deep Consistency Metric Learning Network that leverages metric learning to extract consistent features between single-modal medical images.Specifically,an adaptive feature extractor module is designed to capture lowdimensional semantic features of different input images,which are mapped into a feature embedding subspace.The feature distribution is then similarity-measured to enhance the extraction of semantically consistent features across different input images.By learning these low-dimensional semantic consistency features,the landmark detector is able to acquire more accurate anatomical landmark location information.Comprehensive experiments were conducted on collected cervical vertebra X-ray data sets,and the results demonstrate that CMNet is effective in addressing discrepancies between single-modal medical images and has superior performance compared to the current state-of-the-art medical image landmark detection methods.In addition,the clinical effectiveness of the CMNet method in assisting physicians in cervical spine motion analysis was further validated by measuring cervical spine motion angle parameters.2.Landmark detection of multimodal medical images based on the regularized cycle consistent generative adversarial networkIn cross-domain landmark detection of multimodal medical images,existing crossdomain adaptive methods based on instance generation fail to ensure consistency in the generated image content,which can lead to accuracy issues in subsequent fine-grained image processing tasks such as landmark detection.To address this problem,this paper proposes a deep Regularized Cycle-Consistent Generative Adversarial Network that combines generative adversarial learning and consistency normalization.Firstly,the regularized landmark detection network and the adversarial generation network are jointly trained to generate pseudo images with high content consistency.The landmark detection network is then trained using the true images and the generated pseudo images extended training set.This study uses the public dataset of the multimodal mitral endoscopic surgical suture detection challenge Adapt OR for experimental validation.The results show that Reg Cycle GAN is effective in maintaining the consistency of the image content structure and achieving the best cross-domain surgical suture landmark detection results,compared to existing cross-domain adaptive methods based on instance generation adversarial.3.Multimodal medical image classification based on the deep self-supervised adversarial domain adaptive networkIn the field of unsupervised multimodal cross-domain medical image classification,cross-domain adaptive learning methods don’t use any labeling information from the target domain,resulting in model distribution approximations that are still biased toward the source domain and ultimately fail to yield better classification results on the target domain.This paper proposed an improved deep Self-Supervised Adversarial Domain Adaptive Network to learn semantic consistency features among multimodal images through self-supervised learning and improved adversarial domain adaptive methods,without additional artificial labeling information.SADAN uses a selfsupervised rotating agent task to generate self-rotating labels and enhance the model’s learning of high-dimensional semantic consistency features between multimodal images.A feature-level hybrid module is proposed to overcome the large feature distribution bias between different modalities by learning domain soft labels adversarially with a discriminator.Conduct extensive experiments on the public dataset Ver Se CT and an autonomously constructed X-ray dataset.The experimental results show that,without additional manually labeled target domain information,SADAN method accurately extracts cross-domain semantic consistency features between multimodal data by learning self-rotating labels and domain soft labels,which can better assist physicians in diagnosing osteoporotic vertebral fractures(OVFs)on X-rays.In summary,this paper aims to address the challenges and limitations of current deep learning methods for cross-domain analysis of medical images by overcoming the differences in feature distribution between different medical image domains,and conducts in-depth research on cross-domain medical image recognition,cross-domain semantic consistency feature learning and cross-domain adaptive model construction.These researches have strong theoretical significance and clinical application value.