Deep-Learning-Based Microwave-Induced Thermoacoustic Tomography Applying Realistic Properties of Ultrasound Transducer

Abstract

Microwave-induced thermoacoustic tomography (MITAT) is a noninvasive hybrid modality that has been widely applied in a bunch of biomedical applications. MITAT integrated with cutting-edge deep learning (DL-MITAT) technique is highly promising for tackling many challenging problems that cannot be efficiently solved by traditional methods. However, the previous MITAT or DL-MITAT works rarely consider the realistic properties of the utilized ultrasound transducer, which can significantly degrade the image quality in some application scenarios. To address this issue, we propose a DL-MITAT technique applying realistic properties of ultrasound transducers (referred to as DL-MITAT-PoT), which is very effective for imaging large or long samples. To be specific, the previous related works simply assume an omnidirectional receiving pattern of the transducer. Instead, we take into account the limited receiving angle of a realistic transducer in the DL-MITAT-PoT technique to improve the imaging performance. To implement this technique, we incorporate the receiving properties of a realistic transducer into the training process of the ResU-Net network. We test the trained network by imaging several 2-mm-diameter long blood-vessel-mimicking samples via both simulations and experiments. The obtained imaging results demonstrate that the proposed DL-MITAT-PoT yields much better imaging quality than its counterpart without considering the property of the transducer. We show that the network trained using complicated vessel sample data is endowed with downward compatibility for samples having less complexity. This technique greatly improves the image quality for dealing with large or long samples and has a promising prospect in many biomedical applications.