Research On Visual Cognition Based Medical Image Processing Algorithms And Applications

Recently,the application and achievements of medical image processing in the medical field have been widely recognized.Medical images contain key information on numerous diseases,which can provide important support for disease diagnosis,treatment,and prognosis.Due to the large amount of medical image data,analyzing and processing them requires a lot of time and effort and can lead to errors and omissions.With the continuous advancement of medical image processing technology,these issues have been gradually resolved,and the efficiency of medical image analysis and processing has continued to improve,effectively enhancing the level of medical diagnosis and treatment.Deep learning is a machine learning algorithm that possesses powerful expression and fitting capabilities.Deep learning technology has achieved great success in processing natural language,images,and videos.In addition,it has significant potential in medical image processing.Nevertheless,deep learning still faces several challenges in clinical scenarios.There are several challenges in applying deep learning technology to medical image processing,such as noise and artifacts that can reduce image quality? equipment and acquisition parameters that severely affect the generalization performance of deep learning models? insufficient training data and sample imbalances that can lead to overfitting,and differences between individuals and the complexity of medical images that can also affect model accuracy.Therefore,these issues must be fully considered when applying deep learning technology to medical image processing.To address these challenges,this dissertation was inspired by the mechanisms of brain visual cognition and used various visual attention and search strategies as the methodology to fully borrow from and simulate,giving the deep learning method similar information processing and analysis capabilities to the human brain to more effectively address specific tasks in medical image processing.The main contributions and innovative points of this dissertation are as follows:(1)The first work of this dissertation focused on an improved deep learning model based on the theory of visual-guided search,and it was applied to segment COVID-19 lesions.Characteristics of COVID-19 include common imaging findings,such as groundglass opacities,pulmonary parenchyma,and fibrosis,that are prominently visible in CT images.Quantitative analysis of these lesion features can directly reflect the changes of inflammation in treatment,which is a fundamental factor for evaluating the prognosis of the patient.However,due to the urgency of the COVID-19 epidemic,it is difficult to label large amounts of data for model training.Moreover,the size,distribution,and texture of lesions differ widely,and lung CT images include various lung features that may be misinterpreted as infected areas,further complicating lesion segmentation.Given these challenges,the study created a spatial channel-wised attention network using the theory of guided search.The model used a serial approach to embed spatial and channel attention into its architecture,which resulted in high-quality segmentation performance for COVID-19 lesions.Moreover,the model can create flexible,multi-stage attention maps that present regions of interest in an understandable way,which increases the interpretability of deep learning.Multi-center validation experiments confirmed the model’s excellent ability to generalize and perform well on external datasets.The model was shown to quantitatively assess lung damage and provide effective assistance to radiologists.(2)The second work of this dissertation investigates a semi-supervised one-shot learning method that utilizes visual perceptual organizational rules for segmenting blood vessels in OCTA images of the retina.The vascular system in the retina plays a significant role in determining the severity of systemic,metabolic,and hematological diseases,and can also be used for evaluating the effectiveness of treatment and monitoring disease progression.Manual vessel annotation is a daunting task given the high level of training that medical doctors require and the significant amount of time and effort needed for the annotation process.As a result,implementing it on large-scale,highly dynamic,highresolution data presents a significant challenge.Furthermore,ocular diseases and imaging noise can have a detrimental effect on the clarity and tissue integrity of OCTA images,thereby making vessel segmentation more challenging.This study developed a novel deep learning model and semi-supervised learning method based on visual perceptual organizational rules.The proposed semi-supervised learning method is based on the assumption of covariant rule,which makes the small sample data more dense on the low dimensional manifold,thus improving the convergence of the model with few labeled data.Based on the experimental results,the method proposed in this study exhibits greater segmentation accuracy and generalization performance in one-shot learning tasks when compared to other advanced vascular segmentation algorithms.(3)The third work of this dissertation investigates a deep network model based on dual-channel visual perception and applies it to the classification task of thyroid nodules in ultrasound images.The accurate classification of benign and malignant thyroid nodules is critical in guiding patient treatment and informing surgical decision-making.The objective of this study is to extract local and global feature types and effectively utilize these features to achieve accurate classification.The proposed model is inspired by the dual pathway perception theory of the human visual system and enables multi-task information sharing,thereby facilitating the model to adaptively disentangle the local and global features in the feature space during the optimization process.Experimental results has indicated that this model exhibits superior performance in thyroid nodule classification tasks when compared to other models.Additionally,this study explores the use of artificial intelligence tools in a joint diagnostic model to assist physicians.Experiments have demonstrated that the proposed method enhances the diagnostic accuracy of clinical physicians for thyroid nodules,highlighting its potential clinical utility.In summary,this dissertation focuses on medical images as the research subject and utilizes methods inspired by visual attention to address the specific challenges and limitations of deep learning.It not only attempts to offer new prospects for intelligent information processing of clinical medical problems but also demonstrates the potential of artificial intelligence technology in advancing healthcare.