Image Quality Enhancement with Differentiable Beamforming
Sunday, April 7, 2024
11:50am – 12:15pm
Location: Governor's A - 4th Floor
We present differentiable beamforming (DB), a powerful new approach to image reconstruction that optimizes beamforming quality via autodifferentiation. In DB, the beamformer is expressed as a function of some unknown parameters, e.g., the sound speed throughout the medium, the shape of a flexible transducer, or the position and orientation of a swept transducer. Gradient descent is then used to find the parameters that optimize the focusing quality of the beamformer. With the appropriate choice of focusing criterion, the optimal beamforming parameters coincide with the ground truth. We demonstrate DB in applications of ultrasound autofocusing and sound speed imaging, flexible array shape estimation for wearable applications, and sensorless freehand swept synthetic aperture imaging.