Study On Recognition And Assessment Of Intravascular Ultrasound Plaque Based On Deep Learning Network

Coronary artery disease(CAD)is one of the most deadly diseases in the world.Coronary atherosclerotic plaques are the common causes of coronary heart disease.The IVUS technology uses the interventional catheter to acquire the real-time imaging inside the blood vessel,so as to obtain the anatomical and histological information inside the lumen of the blood vessel,therefore,it has become the new "gold standard" for clinical diagnosis of coronary artery disease.However,the clinical diagnosis of atherosclerotic plaques often depends on the doctor's subjective experience and lacks uniform and objective assessment criteria.This study investigates the plaques recognition and assessment using deep learning network to improve the accuracy and efficiency in clinical diagnosis and treatment.First,intravascular ultrasound data were collected and a total of 735 slices in 60 image sequences from 43 patients were collected.A program was designed to help clinicians to determine the intima region and three common types of plaque such as calcification(240 slices),fiber(219 slices),and lipid plaque(185 slices)as the samples,which were divided into training set(80%)and independent test set(20%).Second,the deep learning segmentation models are constructed to extract the medial and endocardial area where the intravascular plaques occur.The residual structure and the dense cascade structure are introduced to construct three types of deep networks including the U-Net,the Dense-U-Net and the Res-U-Net to fulfill the segmentation task.Third,the deep learning models are adopted to extract three types of plaque and the models are evaluated quantitatively.In media-intima region segmentation,the experimental results show that the best Dice value reached to 85.05%(Res-U-Net),and the highest percentage of areas different(PAD)and Hausdorff distance(HD)are 10.11% and 0.6398 mm.In the plaque identification task,the best Dice values are 0.7078 for calcification plaque(Res-U-Net),0.7145 for fiber plaque(Res-U-Net),and 0.6398 for lique plaque(Dense-U-Net),respectively.Finally,the quantitative assessment of the testing set and plaques burden of individual patient is performed to verify the validity of the proposed model.The minimum lumen area,cross-sectional area and plaque burden are computed to determine if intervention treatment is required.The results show that the plaque identification and assessment model has great potentials for clinical application.