Convolutional Recurrent Neural Network for Bubble Detection in a Portable Continuous Bladder Irrigation Monitor

Continuous bladder irrigation (CBI) is commonly used to prevent urinary problems after prostate or bladder surgery. Nowadays, the irrigation flow rate is regulated manually based on the color (qualitative estimation of the blood concentration) of the drainage fluid. To monitor the blood concentration quantitatively and continuously, we have developed a portable CBI monitor based on the Lambert-Beer law. It measures transmitted light intensity via a camera sensor and deduces the blood concentration. To achieve high reliability, we need to guarantee that the measurement is conducted when there is no air bubble passing through the view of the camera. To detect bubble occurrences, we propose a convolutional recurrent neural network with a sequence of images as input: the convolutional layers extract spatial features from 2D images; the recurrent layers capture temporal features in the image sequence. Our experimental results show that our network has smaller scale and higher accuracy compared with conventional convolutional and recurrent neural networks.