Abstract
Photoacoustic computed tomography (PACT) reconstructs high-resolution images of various chromophores in deep biological tissue. A key to high-quality reconstruction is accurate compensation for the spatially heterogeneous speed of sound (SoS) in tissue. Existing computational methods often estimate or compensate SoS by tuning it directly in the image domain, for example by optimizing sharpness or contrast of reconstructed PA images. However, because the PA signal-to-noise ratio (SNR) decays rapidly with depth due to optical attenuation, such image-domain cues become less informative in deeper regions, limiting SoS accuracy there. Here, we present a dual-modal deep learning framework to correct the heterogeneous SoS via joint processing co-registered PA and ultrasound (US) images. We estimate the spatially varying SoS map from the US image and then fuse the SoS map with the PA image to compute a reduced-aberration photoacoustic image. This method takes advantages of the rich speckle and high SNR in the co-registered US image - and thus can compensate for SoS with high accuracy and efficiency. We tested this method on numerical and tissue-mimicking phantoms, demonstrating cross-domain generalization. In-vivo results demonstrate that incorporation of the predicted SoS maps significantly improved PA image quality, enhancing structural detail and reducing acoustic artifacts. Via fusing the US and PA images, our method produces high-contrast PA images with significantly reduced SoS distortion and artifacts.