Nondestructive ultrasound molecular imaging based on a neural network approach utilizing post-processed ultrasound images

Ultrasound molecular imaging (USMI) utilizes targeted microbubbles (TMBs) for early cancer detection. To improve TMB detectability, differential targeted enhancement (DTE) has been proposed for differentiating molecularly-bound TMBs from unbound TMBs. However, several limitations to DTE, such as non-real-time imaging, the necessary destruction of TMBs, and the limited field-of-view prevents it from being clinically useful. Thus, the feasibility of nondestructive USMI through neural network approach has been suggested.Here, we propose a neural network based USMI approach based on post-processed ultrasound images as inputs. We implemented a simple 4-layer convolutional neural network (CNN) to achieve nondestructive USMI. The network approach outputs the pixel-wise probability of TMB presence. We performed an in vivo animal study to evaluate the proposed nondestructive USMI mode. 14 transgenic mice with spontaneous breast cancer were imaged. Ultrasound videos were recorded before and after MB injection and destruction, which were used to generate 2D images for network training and testing. Our data was augmented by injecting 4 types of MBs, extracting multiple DTE frames from the video segments, and capturing multiple random patches from the images; resulting in 15,000 patches for the network training and testing. The testing data were used to evaluate the network performance.MB detection within lesions showed a strong correlation with DTE: correlation coefficient (R) = 0.978, p-value < 0.001.The continuous dice coefficient (CDC) was used to evaluate localized MB detectability. In cases where the MB detection percentages were greater than 20%, a high CDC of 0.863 was achieved. DTE and network prediction images were compared; network prediction images highlighted MBs comparable to DTE and, moreover, showed less false positive detection for tumor boundaries compared to DTE. We anticipate that utilizing widely available postprocessed ultrasound images can contribute to enable the development of clinically available nondestructive USMI.