Data synthesis for learning speckle patterns from simulated ultrasound signals

Sammanfattning

Ultrasound imaging is widely used in healthcare as a safe and cost-effective tool for diagnosis.The field of quantitative ultrasound (QUS) is gaining increased interest for its potential toenable more accurate diagnoses through biomarkers extracted from ultrasound data. One suchbiomarker is tissue scatterer density (scatterers per mm2), which can be used to differentiatehealthy from diseased tissue and has applications in detecting conditions such as cancer andliver disease. Knowledge of the imaging system’s resolution is also valuable, both for qualitycontrol and for facilitating the prediction of scatterer properties.Most existing methods for extracting quantitative parameters rely on radio-frequency or envelopedata, and using deep learning on B-mode images for this purpose remains unexplored. Trainingdeep learning models on real ultrasound data is challenging due to the difficulty of obtaininglarge datasets with known scatterer properties. This study investigates the use of multitaskdeep learning to predict scatterer density and image resolution directly from ultrasound Bmodeimages. To overcome the problem with data availability, a large dataset of image patcheswas generated using a linear simulation method, allowing full control over scatterer densityand system parameters. A multitask model inspired by U-Net was trained exclusively on thissynthetic data and evaluated on both simulated test images and real B-mode images acquiredfrom two phantoms.The results demonstrate that multitask learning, predicting resolution alongside scatterer density,improves scatterer density estimation. The best performing model achieved a mean absoluteerror below 5% on both scatterer density and resolution prediction. Despite being trained onlyon simulated data, the model seemed to generalize reasonably well to real ultrasound images,showing consistent predictions across varying imaging parameters. These findings suggest thatdeep learning models trained on synthetic B-mode data is a feasible approach for quantitativeultrasound, potentially lowering the barrier to clinical adoption of QUS techniques.