Inlier Detection in Thermal Sensitive Images

Abstract

Image guidance of medical procedures may use thermal images to monitor a treatment. Analysis of the thermal images by the physician may be time consuming and confusing because the thermal image includes multiple outliers. We present a novel inlier detection method for thermal images that results in reliable thermal information to support medical decision making. Outliers in thermal images are particularly challenging to detect using conventional methods, because they are significantly more abundant than inliers and, like inliers, they may be temporally consistent. Our inlier detection method is physically-based: it is motivated by the fact that heat propagation in soft tissues can be modeled using the bio-heat equation. Pixels are classified as inliers only if the temperature pattern in a spatial and temporal neighborhood strongly correlates with the physical model. For improved robustness, the correlation process includes a 2D filter in the spatial domain and a 3D filter in both spatial and temporal domains. Experiments with real data have shown that our method produces results that agree with annotations provided by human experts even in outlier-laden images. Our results show inliers can be detected leaving true heat pixels for the physician to observe, while not overloading him with the need to analyze outliers. The technique has been integrated in a true clinical environment and is being used to aid physicians in analysis of thermal images