Convolutional Neural Network for Photoacoustic Imaging Segmentation Applied in Human Study

Abstract

Photoacoustic microscopy (PAM) is a hybrid modality of biological imaging (skin layers, normal or damaged blood vessels, and tumors) technique based on the photoacoustic effect generated by Lasers. To easily visualize and analyze the significant regions, now a day’s segmentation techniques are used more frequently. This technique is directly associated with the more accurate diagnosis of human diseases such as diabetes, dermatology, oncology, and neurology. However, many studies using classical segmentation methods have encountered several limitations in the segmentation of photoacoustic signals obtained from the targeted tissues, mostly associated with noises, low image depth, and poor image resolution. This thesis proposes an U2-Net neural network to segment each pixel in the PA image individually. This method is based on the original B-scan dataset and label dataset for training, followed by making predictions to generate segment regions. The study was performed on human hands to segment blood vessels and skin layers, whereas mice models were used to segment the tumor and healthy blood vessels. The following experimental results show that the U2-net model could achieve a good PAI segmentation performance for all the experimental (human hand) datasets, where the average pixel accuracy reached ~94%. Therefore, this method could be a promising easy-to-process technique for PAM imaging with enhanced quality for clinical-level photoacoustic imaging diagnosis application