Automated liver tumor detection in abdominal ultrasonography with a modified faster region-based convolutional neural networks (Faster R-CNN) architecture

BACKGROUND: Although diagnostic ultrasound can non-invasively capture the image of abdominal viscera, diagnosis of the continuous ultrasound liver images to detect a liver tumor effectively and to determine whether the detected is benign or malignant is nontrivial. In order to minimize the gaps in diagnostic accuracy depending on doctor’s proficiency, we built an automated system to support the ultrasonography of liver tumors by employing deep learning technologies. METHODS: We constructed a neural network model for the automated detection of tumor tissues and blood vessels from the sequential liver ultrasound images. Faster region-based convolutional neural networks (Faster R-CNN) is employed as a base model for the object detection, which can output the detection results in 4 frames per second and enable the system to be particularly suitable for the real time ultrasonography. Moreover, we proposed a new neural network architecture feeding both the current and previous images into Faster R-CNN. For training the models, intraoperative ultrasound images obtained from one hepatocellular carcinoma (HCC) patient were used. The obtained image was a multifaceted observation of the liver and includes one HCC and some blood vessels. We labeled 91 images with the help of a liver specialist. We compared the tumor detection performance of the plain Faster R-CNN model with that of the proposed model. RESULTS: We find that both the models performed well in detecting HCC and blood vessels, after training with 400 epochs using Adam. However, the mean precision of our model reaches 0.549, which is 0.019 better than that of the plain Faster R-CNN, and the mean sensitivity of our model about HCC reaches 0.623±0.385 for 30 scenes of sequential liver ultrasound images, which is also 0.146 better than that of the plain Faster R-CNN model. CONCLUSIONS: The comparison between the proposed model and the plain Faster R-CNN model shows that we achieved better accuracy in tumor detection, in terms of the mean precision as well as the mean sensitivity, with the proposed model.