Improved Imaging and Quantitative Identification of acoustic tissue properties with dual band ultrasound pulse complexes

Tissue characterization involves classifying tissue types based on some distin- guishable property of the tissue. There are a wide variety of applications where distinguishing one type of material from another provides benefit and insight. One such application is in medicine where the identification of both pathological and healthy tissue can provide added value in diagnosis and further treatment planning. Another example is in the meat industry where tissue composition is strong pre- dictor of meat quality and eating experience. In both applications there is a strong need for non-invasive characterization methods. In the medical sector, invasive methods can cause discomfort or even harm, and are often hard to conduct as nav- igation can be challenging. Correspondingly, in the meat industry, unnecessary harm or discomfort should be minimized and characterization done post-mortem limits genetic progress as selection for breeding must be determined through sib- lings. Performing in vivo imaging is typically done with three main modalities; Ul- trasound (US), X-ray Computed Tomography (CT) and Magnetic Resonance Ima- ging (MRI). Out of these three, ultrasound is the most portable and inexpensive method, albeit with the poorest overall image quality. In recent years, the field of medical imaging analysis has undergone a revolu- tion. With the increasing popularity of artificial intelligence and machine learning, more and more of these techniques are being utilized in automatic image analysis. Deep learning has proven to be especially successful, a data-driven approach able to distinguish complex features in data. This work investigates the potential for using deep learning in tissue charac- terization applications, and at what point the performance of the models becomes limited by the underlying data. It is found that for segmentation of CT volumes, deep learning achieves a high performance allowing automatic labeling of volumes and enabling robust Atlas segmentation. The same approach is used in an attempt to quantify fat content from pulse-echo ultrasound images, producing results com- parable to state-of-the-art approaches. However, for high fat content the results deteriorate, producing a limit on the range of applicability due to decreasing image quality. To produce robust tissue characterization with ultrasound, investigation of other acoustic parameters is necessary. In particular it is found that the nonlinear bulk elasticity of soft tissues has a large variation compared to other acoustic paramet- ers. In addition, the parameter is especially sensitive to fat content. Further, it is shown that this parameter can be measured using a dual frequency approach. The presented approach has the characteristic hallmarks of an approach in which deep learning can be successful, and the potentially increased data quality can provide a more robust tissue characterization method. However, as deep learning is data-driven, the main challenge will be the procurement of realistic training data, consequently becoming an interesting area of further study.

-

Ultralyd hastighet, forplantning og spredning bestemmes av volum kompressibiliteten til vevet. Denne er dominert av vannet i vevet, slik at ultralyd hastigheten c ? 1500 m/s, nær den samme som i vann. I tillegg har en i bløtt vev skjærbølger, også kalt deformasjons bølger, som ikke skaper volum kompresjon. Disse har lav bølgehastighet ~ 1 - 10 m/s som varierer sterkt med struktur av store molekyler i vevet. Skjærbølgehastigheten er derfor nyttig til å differensiere mellom ulike typer vev, spesielt evaluere vevstyper som angitt i hovedmålet. Ved SURF imaging undertrykkes multiple støy i bildet sterkt, og en er i stand til å skille mellom lineær og ulineær spredning. I tillegg estimerer man en ulineær elastisitets parameter som kan benyttes til å skille vevstyper, hvor for muskelvev og for animalsk fett. For flerumettede fettsyrer forventes denne parameteren å være større, og man forventer en glidende overgang mellom fett og muskelverdier avhengig av volumprosenten av fett i muskelen. Redusert støy i bildene forbedrer også bestemmelsen av skjærbølge-stivhet i vevet, slik at man har mulighet for å trekke ut to kvantitative parametre som avhenger av detaljer i vevsstrukturen. Ut fra forskningen ved ISB er det med hjelp fra NTNU-TTO etablert et nytt firma, SURF Technology AS. Firmae har hatt støtte fra NFR-FORNY2020 for å utvikle et demonstrator instrument for SURF imaging. Man planlegger å bygge flere instrumenter utover i 2016. Det gjenstår flere interessante utfordringer innen signalbehandling, spesielt kvanttativ estimering av samt skjærbølge stivhet. Det kan også være behov for nye design av to-frekvensprober for de spesielle anvendelsene for dyr og fisk. Nye metoder implementeres som SW med beskjedne HW endringer i scanneren. PhD prosjektet består derfor av teoretisk utvikling av metodene med SW implementering i scanner, samt utprøvning av nye metoder i laboratorium og i dyr. Prosjektet har delmål som angitt i neste kapittel.