Research On The Quality Assessment Method Of Ophthalmic Artery MECT Image Based On Traditional Features And Siamese Network

Adaptive statistical iterative reconstruction V(ASi R-V)technology of Multi-energy computed tomography(MECT)is widely used in clinical imaging.Different energy values and ASi R-V mixing weights in the imaging process have different degrees of influence on image quality.Therefore,accurate quality assessment of ophthalmic artery MECT images is crucial.Traditional image quality assessment methods include both subjective and objective perspectives.Due to the complexity and particularity of medical images,subjective assessments are time-consuming and labor-intensive,also,there are subjective differences among observers,what’s more,no unified assessment of objective and non-referenced quality assessment method of medical images.Methods to evaluate image quality from an objective perspective using deep learning are widely used,but it is difficult to obtain reference images from ophthalmic artery MECT images,resulting in the lack of labeled samples for training in deep learning-based evaluation methods.This paper studies the noreference quality evaluation of ophthalmic artery MECT images at the two levels of traditional features and deep learning to solve the above problems.The main work is as follows:First,the traditional feature-based ophthalmic artery MECT image quality assessment method.In this paper,two methods based on sub-image structural similarity(method 1)and based on feature domain contrast-to-noise ratio(method 2)are used to evaluate the quality of ophthalmic artery MECT images without reference: Method 1 divides the ophthalmic artery cross-sectional slice of the ophthalmic artery MECT image into sub-images,decompose the features of the original image into different sub-images,and pay attention to the changes in the structure of the decomposed images from different directions;Method 2extracts a series of texture features of the ophthalmic artery region of interest(foreground)and surrounding tissue(background)and perform feature screening.According to the idea of contrast-to-noise ratio,the foreground and background are compared in the feature domain.According to the idea of contrast-to-noise ratio,the foreground and background are compared in the feature domain.The root mean square error(RMSE)of the evaluation results of the two methods is 1.212 and 0.938,respectively,the linear correlation coefficient(LCC)is 0.672 and 0.741,respectively,and the Spearman rank correlation coefficient(SROCC)is 0.660 and 0.735,respectively,proved that they both have a high correlation with the subjective evaluation of physicians,and the method based on the contrast-to-noise ratio of the feature domain has a better evaluation effect.Both methods do not require reference images,which realizes the application of traditional image quality assessment methods in medical images,and is in good agreement with the visual focus of radiologists when diagnosing diseases,and has certain clinical auxiliary diagnostic value.Second,an ophthalmic artery MECT image quality assessment method based on Siamese network.In this paper,a model based on Siamese network is proposed for no-reference quality assessment of ophthalmic artery MECT images: Construct a paired dataset of image patches of the ophthalmic artery blood vessel and the surrounding background,use the Siamese network to extract the contrast features between the two,and take the difference between the two image patches as the target to generate a representation of the similarity between the two images.The scalar is used as the prediction of the image quality score,and the overall features of the ophthalmic artery cross-section slice images are extracted through the overall feature extraction module,and the contrast features are used to guide the model training.The RMSE of the evaluation results of this method is 0.882,which is 0.330 and0.056 lower than the two methods based on sub-image similarity and feature domain contrast-to-noise ratio,respectively.The LCC is 0.829,which is improved by 0.157 and0.088.The SROCC is 0.814,an improvement of 0.154 and 0.079.It is demonstrated that the Siamese network-based method is higher correlated with the subjective assessment of physicians.This method generates paired data sets that do not use the physician’s subjective assessment as a reference label,and makes full use of the contrast information between the ophthalmic artery blood vessel and the background image of the surrounding tissue and the overall information of the ophthalmic artery cross-sectional slice,which provides a new idea for the quality assessment of ophthalmic artery MECT images.