Séminaire CREATIS - Improving diagnosis by ultrafast imaging and artificial intelligence

Chris L. de Korte, Radbout University

Résumé au format pdf + biographie de l'orateur

Recent developments in computer processing and hardware and the introduction of deep learning are revolutionizing ultrasound imaging. The diagnostic performance of high-end ultrasound systems is improved by these developments and allows functional imaging in 3D. Developments in transducer technology facilitates 3D imaging as well as ultra-broadband and high frequency ultrasound imaging. Additionally, cheap transducers connected to laptops or smart phones allow basic ultrasound scanning for a very affordable price.

Quantification of plaque compositions and the remaining velocity profile is crucial for adequate therapy in atherosclerotic disease. We developed high frame rate methods for characterization of the arterial wall, plaques and flow. First, we developed a compound strain imaging technique in which the carotid is imaged at a combination of large beam steered angles to fully quantify the 2D strain vector. The presence of high strain regions is considered to be related to rupture prone spots of the plaque. Currently, we are transferring this technique from 2D to 3D by combining compound strain imaging with high frame rate imaging. Additionally, push wave elastography for transverse vessel cross-sections is being developed. In this technique, a shear-like wave is generated by an ultrasound push and its circumferential propagation is captured by high frame rate imaging. Since the propagation behavior of the shear-like wave is directly related to the modulus of the arterial wall and plaque it provides information on the plaque composition. Finally, we utilized compound imaging to quantify the 2D blood velocity vector. By acquiring over 10.000 fps at two beam steered angles, the full 2D velocity vector can be determined. This technology has proven to provide superior velocity estimates over the full range of slow to high velocity values. Compound strain and flow imaging were validated using a realistic bifurcation phantom and a small-scale study in patients and volunteers.

Ultrasound imaging is used for breast cancer detection in women with dense breast in which mammography shows a reduced sensitivity by the higher amount of glandular tissue. Since hand-held ultrasound is operator dependent, the automated breast volume scanner (ABVS) was introduced that consists of a linear array transducer that is translated motor-controlled over the breast while collecting ultrasound data to reconstruct a volumetric breast image. Although clinical studies show high sensitivity, clinicians report high recall-rates due to the detection of many lesions of uncertain malignant potential. Compared to benign lesions, malignant lesions are often stiffer, and more grown into the surrounding tissue (firmly bonded) resulting in decreased strains inside, and shear strain around the lesion respectively. We extended 2D strain imaging to full 3D strain imaging using an Automated Breast Volume Scanning (ABVS) system by incorporating strain imaging in combination with ultrafast plane wave imaging. Validation studies in breast phantoms and in patients (n=43) reveals the potential of this technique to identify and classify lesions.
Low cost ultrasound devices in combination with automated analysis will increase the application dramatically. We developed a prenatal scanner based on low cost ultrasound in combination with a standardized protocol and deep learning for low resource countries. The scanning procedure does not require knowledge and can be taught to heath care workers in 2 hours. The system automatically determines the presence of single pregnancy or twin, the position (breach or normal) of the fetus and the age and was evaluated in 240 women in Ethiopia.