TY - JOUR
T1 - Faster super-resolution ultrasound imaging with a deep learning model for tissue decluttering and contrast agent localization
AU - Brown, Katherine G.
AU - Waggener, Scott Chase
AU - Redfern, Arthur David
AU - Hoyt, Kenneth
N1 - Publisher Copyright:
© 2021 IOP Publishing Ltd
PY - 2021/11
Y1 - 2021/11
N2 - Super-resolutionultrasound(SR-US)imagingallowsvisualizationofmicrovascularstructuresassmallas tensofmicrometersindiameter.However,useintheclinicalsettinghasbeenimpededinpartbyultrasound (US)acquisitiontimesexceedingabreath-holdandbytheneedforextensiveofflinecomputation.Deep learningtechniqueshavebeenshowntobeeffectiveinmodelingthetwomorecomputationallyintensive stepsofmicrobubble(MB)contrastagentdetectionandlocalization.Performancegainsbydeepnetworks overconventionalmethodsaremorethantwoordersofmagnitudeandinadditionthenetworkscanlocalize overlappingMBs.TheabilitytoseparateoverlappingMBsallowsuseofhighercontrastagentconcentrations andreducesUSimageacquisitiontime.Hereinweproposeafullyconvolutionalneuralnetwork(CNN) architecturetoperformtheoperationsofMBdetectionaswellaslocalizationinasinglemodel.Termed SRUSnet,thenetworkisbasedontheMobileNetV3architecturemodifiedfor3-Dinputdata,minimal convergencetime,andhigh-resolutiondataoutputusingaflexibleregressionhead.Also,weproposeto combinelinearB-modeUSimagingandnonlinearcontrastpulsesequencing(CPS)whichhasbeenshown toincreaseMBdetectionandfurtherreducetheUSimageacquisitiontime.Thenetworkwastrainedwithin silicodataandtestedoninvitrodatafromatissue-mimickingflowphantom,andoninvivodatafromtherat hindlimb(N = 3).ImageswerecollectedwithaprogrammableUSsystem(Vantage256,VerasonicsInc., Kirkland,WA)usinganL11–4vlineararraytransducer.Thenetworkexceeded99.9%detectionaccuracyon insilicodata.Theaveragelocalizationaccuracywassmallerthantheresolutionofapixel(i.e.l/8).The averageprocessingtimeonaNvidiaGeForce2080TiGPUwas64.5msfora128 × 128-pixelimage.
AB - Super-resolutionultrasound(SR-US)imagingallowsvisualizationofmicrovascularstructuresassmallas tensofmicrometersindiameter.However,useintheclinicalsettinghasbeenimpededinpartbyultrasound (US)acquisitiontimesexceedingabreath-holdandbytheneedforextensiveofflinecomputation.Deep learningtechniqueshavebeenshowntobeeffectiveinmodelingthetwomorecomputationallyintensive stepsofmicrobubble(MB)contrastagentdetectionandlocalization.Performancegainsbydeepnetworks overconventionalmethodsaremorethantwoordersofmagnitudeandinadditionthenetworkscanlocalize overlappingMBs.TheabilitytoseparateoverlappingMBsallowsuseofhighercontrastagentconcentrations andreducesUSimageacquisitiontime.Hereinweproposeafullyconvolutionalneuralnetwork(CNN) architecturetoperformtheoperationsofMBdetectionaswellaslocalizationinasinglemodel.Termed SRUSnet,thenetworkisbasedontheMobileNetV3architecturemodifiedfor3-Dinputdata,minimal convergencetime,andhigh-resolutiondataoutputusingaflexibleregressionhead.Also,weproposeto combinelinearB-modeUSimagingandnonlinearcontrastpulsesequencing(CPS)whichhasbeenshown toincreaseMBdetectionandfurtherreducetheUSimageacquisitiontime.Thenetworkwastrainedwithin silicodataandtestedoninvitrodatafromatissue-mimickingflowphantom,andoninvivodatafromtherat hindlimb(N = 3).ImageswerecollectedwithaprogrammableUSsystem(Vantage256,VerasonicsInc., Kirkland,WA)usinganL11–4vlineararraytransducer.Thenetworkexceeded99.9%detectionaccuracyon insilicodata.Theaveragelocalizationaccuracywassmallerthantheresolutionofapixel(i.e.l/8).The averageprocessingtimeonaNvidiaGeForce2080TiGPUwas64.5msfora128 × 128-pixelimage.
KW - Contrast-enhanced ultrasound
KW - Deep learning
KW - Microbubbles
KW - Plane-waves
KW - Super-resolution ultrasound
UR - http://www.scopus.com/inward/record.url?scp=85119423031&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85119423031&partnerID=8YFLogxK
U2 - 10.1088/2057-1976/ac2f71
DO - 10.1088/2057-1976/ac2f71
M3 - Article
C2 - 34644679
AN - SCOPUS:85119423031
SN - 2057-1976
VL - 7
JO - Biomedical Physics and Engineering Express
JF - Biomedical Physics and Engineering Express
IS - 6
M1 - 065035
ER -