Data-Driven Learning to Detect Characteristic Kinetics in Ultrasound Images of Arthritis

摘要

Contrast Enhanced Ultrasound (CEUS) is a sensitive imaging technique to assess synovial vascularization and perfusion, allowing a pixel-wise perfusion quantification that can be used to distinguish different forms of disease and help their early detection. However, the high dimensionality of the perfusion parameter space prevents an easy understanding of the underlying pathological changes in the synovia. In order extract relevant clinical information, we present a data-driven method to identify the perfusions patterns characterizing the different types of arthritis, exploiting a sparse representation obtained from a dictionary of basis signals learned from the data. For each CEUS examination, a first clustering step was performed to reduce data redundancy. Then a sparse dictionary was learnt from the centroids. The perfusion time-curves were represented as a sparse linear combination of the basis signals, estimating the coefficients via a LASSO algorithm. With this representation, we were able to characterize each pathology through a small number of predominant kinetics. By using sparse representation of CEUS signals and data-driven dictionary learning techniques we were able to differentiate the specific kinetics patterns in different type of arthritis, suggesting the possibility of personalizing the description of each patient's type of arthritis in terms of relative frequency of the detected patterns. Interestingly, we also found that rheumatoid and psoriatic arthritis share some common perfusion behaviors.