Deep compressed sensing for characterizing inflammation severity with microultrasound

With histological information on inflammation status as the ground truth, deep learning methods can be used as a classifier to distinguish different stages of bowel inflammation based on microultrasound (µUS) B-scan images. However, it is extremely time consuming and animal usage is high to obtain a balanced data set for every stage of inflammation. In this study, we describe a deep compressed sensing method to increase the number of B-scan images for inflammation studies without use of additional animals. In this way, training data can be quickly augmented. The fidelity of the synthesized data is evaluated using both qualitative and quantitative methods. We find that the synthetic data have high structural similarity when compared with original B-scan images. Further evaluation, such as finding the correlation of µUS and microscopy images and calculating attenuation coefficient, will be investigated in future to provide better understanding.