Multi-Task Learning for Simultaneous Speed-of-Sound Mapping and Image Reconstruction Using Non-Contact Thermoacoustics

Multi-modal imaging via thermoacoustic (TA) approaches provides contrast mechanisms differing from conventional ultrasound (US) imaging - opening up new applications like non-invasive, non-contact below-ground sensing. Due to the high correlation between soil moisture content and speed-of-sound (SoS), knowledge about the SoS in soil can be utilized to improve below-ground image reconstruction and soil moisture mapping at depth. In this work, we present multi-task deep learning networks to accurately predict arbitrarily varying SoS distributions in soil while concurrently reconstructing high-fidelity TA images of root structures. We deploy multi-task U-Net based fully convolutional neural networks trained using US data generated through TA simulations on a wheat root dataset. A multi-input, multi-output architecture performed best - achieving the highest root image contrast-to-noise ratio and lowest SoS mean absolute error.