Research On Key Technology For Multi-View Medical Image Anomaly Detection

Medical imaging is a procedure in the radiology or nuclear medicine department,and its rich imaging modalities are widely used in modern medical diagnosis.For medical imaging data,multiple views indicate multiple different ways or different angles of the depiction of the same study object.Doctors’ diagnosis of patients often requires analysis and comparison with the help of multiple views.Therefore,multi-view medical images have changed from auxiliary examination tools to the most important clinical diagnosis and differential diagnosis basis of modern medicine.In addition,most of the traditional methods for medical image analysis rely on manual features,which require relevant experts to spend a lot of time and effort to mark regions of interest in combination with their a priori knowledge.In practical clinical applications,a large number of disease classes and the low prevalence of some diseases lead to problems such as imbalance and modeling difficulties in the available data.In contrast,anomaly detection can train models on most classes of samples and detect a few classes as anomalous samples,which can solve the problems of medical image data.Therefore,the anomaly detection approach is gradually receiving more and more attention in the data processing of medical images.After reviewing the literature,we found that the existing anomaly detection algorithms often ignore the characteristics of clinical and medical images themselves.The anomaly detection of medical images usually involves three aspects: detection,segmentation and diagnosis.Therefore,these algorithms still need to be improved in acquiring distinguishable features of medical images.In summary,how to obtain distinguishable features for medical images with special imaging theoretical and application problems are worth studying to improve the accuracy of the algorithms for abnormality detection in the case of small and uneven sample sizes.In view of the above-mentioned characteristics and problems of data and models in medical image anomaly detection,we design the following algorithms to improve the accuracy:(1)Brain tumor abnormality detection plays a crucial role in the field of computeraided diagnosis.However,brain tumor data is scarce and difficult to classify.Unsupervised methods can reduce the huge labeling cost and can be applied to brain tumor abnormality detection given only normal brain images during training.However,existing unsupervised methods only determine whether a sample is abnormal or not at the image level and cannot learn discriminative features effectively.To address the above problems,a novel brain tumor abnormality detection method is proposed,which obtains discriminative features by adversarial learning based on the regularization of latent space features.Comprehensive experiments on Bra TS,HCP,MNIST,and CIFAR-10 datasets evaluate the effectiveness of this paper,and its performance exceeds that of state-of-the-art brain tumor detection methods.(2)Self-supervised learning has the potential to be effective at capturing generic knowledge about various concepts,which will be helpful for many downstream image analysis tasks.This paper suggests a novel self-supervised learning approach that takes into account the unique multiple imaging modalities of medical images.Additionally,to make learning feature representations for downstream tasks easier,this paper uses a multimodal fusion classification task as a pre-training task to learn a modality-independent feature embedding by confusing image modalities at the data level.The refined learned features can be used for a variety of downstream tasks.Finally,the learned representations are fine-tuned before being applied to the downstream multimodal medical image segmentation task.Extensive testing demonstrates that the algorithm suggested in this paper outperforms cutting-edge techniques on the Bra TS 2019 and CHAOS datasets.(3)U-networks have limitations in terms of global(long-range)contextual interactions and preservation of edge details.In contrast,the Transformer module has an excellent ability to capture long-range dependencies by utilizing a self-attentive mechanism,it still suffers from high computational complexity and spatial complexity when dealing with high-resolution 3D feature maps.To solve the above problems,this paper proposes an efficient Transformer-based U-model.In addition,this paper also proposes a self-distillation unit,which can extract fine-grained details from the skip connections of the encoder while learning global semantic information and local spatial detail features.To further preserve the edge and detailed information,a multi-scale fusion block is first proposed in this paper.Extensive experiments on Bra TS 2019 and CHAOS datasets show that the performance of the algorithm in this paper outperforms the state-of-the-art methods.(4)Functional brain connectivity derived from resting-state functional magnetic resonance imaging has been widely used to study neuropsychiatric disorders such as autism.However,existing studies typically suffer from the following problems:(i)Different scanners or multiple site study populations can cause significant data heterogeneity.(ii)There are millions of voxels in each f MRI scan,while the number of training samples is very limited(tens or hundreds).(iii)Poor interpretability,which hinders the identification of reproducible disease biomarkers.In this work,we propose a multi-site clustering and nested feature extraction method for autism detection based on resting-state functional magnetic resonance imaging.We divided the multisite training data into an autism group and a healthy control group.To model the inter-site heterogeneity in each category,we used a similarity-driven multi-view linear reconstruction model to learn potential representations and perform clustering within each group.Then,we designed a nested singular value decomposition method to mitigate the inter-site heterogeneity,followed by a linear support vector machine for the detection of autism.In the experimental section,this paper demonstrates the effectiveness of the framework and shows superiority over several state-of-the-art methods for autism detection.