Compounding Nakagami Parameter Ratio Imaging and Deep Learning Approach with Contrast-Enhanced Ultrasound for Tissue Lesion Assessment

Abstract

Ultrasound imaging has been widely used to assess the morphological changes of tissue lesions however morphological related changes usually lack of disease specificity Consequently it remains difficult for the majority of ultrasound B-mode imaging to precisely diagnose a specific tissue lesion since the morphological changes of tissues are usually non-disease-specific On the other hand microvascular changes can accurately reflect the severity and recovery of tissue lesions To further alleviate these hurdles the administrated ultrasound contrast agents (UCAs) in the bloodstream allow the acquisition of contrast-enhanced ultrasound (CEUS) imaging over a certain duration to estimate the tissue perfusion Nevertheless the frequency of most of the diagnostic ultrasounds is less than 10 MHz and that tends to result in a short-of-sufficient spatial resolution sampling rate tissue clutter tolerance to measure the blood flow or perfusion in the capillary beds Certainly it is straightforward to increase the resolution of the ultrasound image by the increase of employed ultrasound frequency Nevertheless the increase of ultrasound frequency tends to unavoidably increase the acoustic attenuation and then decrease the depth and contrast of the image greatly Therefore it certainly is desirable to further explore and develop alternate diagnostics for better assessing the states of lesions or treatment effect covering the microvascular changes in local tissue To further alleviate these issues the present study developed a contrast-specific ultrasound imaging system using the non-linear and pressure-dependence characteristics of UCAs To explore the effects of backscattering properties of UCAs the contrast-specific ultrasound imaging system was equipped with three different ultrasonic transducers which were developed for measuring the UCAs suspensions Various pressure amplitudes of transmitted ultrasound ranging from 0 3 to 1 2 MPa corresponding to each ultrasound frequency were also adjusted The tendency of the Nakagami parameter as a function of ultrasound frequency was opposite to that of backscattered power and that the Nakagami parameter of UCAs decreased from 0 92 ± 0 05 to 0 80 ± 0 03 as the driving frequencies increase from 3 to 7 MHz As the UCAs suspensions were insonified at 7 MHz the Nakagami parameter was dramatically decreased from 0 81 ± 0 03 to 0 73 ± 0 02 with the increase of incident acoustic pressure from 0 3 to 1 2 MPa These results indicate the Nakagami parameter can effectively reflect the frequency and pressure changes on UCAs Specifically the perfusion parameters estimated from the ultrasound time-intensity curve (TIC) and statistics-based time-Nakagami parameter curve (TNC) approaches were found able to quantify the perfusion Nevertheless due to insufficient tolerance on tissue clutters and subresolvable effects Nakagami parameter-based approaches remain short of reproducibility and stability Therefore in this study the window-modulated compounding (WMC) Nakagami parameter ratio imaging was proposed to alleviate these effects by taking the ratio of WMC Nakagami parameters corresponding to the incidence of two different acoustic pressures from an employed transducer The time-Nakagami parameter ratio curve (TNRC) approach was also developed to estimate perfusion parameters The verification of contrast-specific system and WMC Nakagami parameter ratio approach were performed from flow phantom and animal subjects administrated with a bolus of UCAs The TNRC approach demonstrated better sensitivity and tolerance of tissue clutters than those of TIC and TNC The fusion image with the WMC Nakagami parameter ratio and B-mode images indicated that both the tissue structures and perfusion properties of ultrasound contrast agents may be better discerned To extensively explore the contrast-specific ultrasound imaging system for noninvasively imaging and perfusion evaluation of tissue lesions and alleviate the effects of subjective and system factors In the present study efforts were made aiming to further improve the detection and classification of muscle injury with ultrasound compound imaging that fused quantitative ultrasound and perfusion parameters Animal experiments were performed from a total of 12 rats where the contusion injury in response to a certain impact was made on their gastrocnemius muscle Each measurement was carried out using the contrast-specific ultrasound imaging system that covered the recovery phases of contusion for three weeks to obtain CEUS images and those just mentioned parameters The muscle recovery phases were classified by designating the gastrocnemius muscle injury level (GIL) of 0 to the healthy stage that corresponds to the uncontused tissue; those of 1 and 2 to the destruction and repair and remodeling phases associated with the certain muscle recovery phases Subsequently three conventional machine learning approaches including naive Bayes (NB) support vector machine (SVM) and artificial neural network (ANN) were employed to classify GILs utilizing quantitative ultrasound and perfusion parameters Moreover AlexNet GoogleNet and VGG-19 were adopted for training to extract the feature maps of those acquired CEUS images for automatically classifying the GIL with a deep convolutional neural network The classification with NB SVM ANN AlexNet GoogleNet and VGG-19 resulted in accuracies of 57 7 ± 8 54 62 22 ± 1 37 66 00 ± 1 39 95 31 ± 0 33% 99 36 ± 0 32% and 99 63 ± 0 35% These results indicated that the CEUS in conjunction with WMC Nakagami parameter ratio imaging has the ability for assessing tissue lesions Moreover the results also demonstrate that TNRC is able to reduce the effect of tissue clutters from injured muscle fibers and scar tissues thus increase the reproducibility for assessing microcirculation In addition this suggests deep learning approaches combined with the CEUS images are of potential to sensitively detect and classify the muscle injury

Date of Award	2021
Original language	English
Supervisor	Shyh-Hau Wang (Supervisor)