Detection and Segmentation of Free Blood in FAST Exam Ultrasound Images
MetadataShow full item record
It is often a diagnostic challenge for emergency physicians and trauma surgeons to detect major internal injuries of trauma patients. Bedside ultrasonography is a fast, accurate, cost effective way to detect life-threatening internal injuries and can be performed in unstable patients. Despite the growing popularity of ultrasound imaging in trauma cases, the automatic segmentation of organs and the detection of free blood remains a real challenge. Performing this life-saving exam in an automatic way would be a significant improvement as it could decrease the time for manual interpretation of the screenings, the subjectivity of these interpretations and also decrease the dependency on experienced emergency physicians. In our study, we find solutions for this real life problem by computer vision. To our knowledge, this problem has not been addressed before. We review the related works and use the state of the art methods in segmentation of ultrasound images as a starting point. We adapt the common approaches and investigate their application on our problem of detecting and segmenting free blood in FAST exam ultrasound images. We evaluate their results and suggest modifications and define new possible features that could be used to characterize tissues and free blood in the images. We handle the problem of denoising of ultrasound images using non local means method. Then we investigate the application of multiphase Chan-Vese active contours without edges method and clustering with textons on our dataset. Then, we improve the study with textons by suggesting using the distribution of textons as a feature and labeling parts of the image with the help of a naive Bayes classifier trained with these features. Finally, we propose a region growing approach based on the textons, in other words using the similarity criterion of filter responses. To our understanding, our study takes a real life problem and addresses it as a new computer vision problem. It investigates the application of state of the art methods on the problem and also suggests modifications and new approaches to solve this problem. With its promising results, our study could be used as a benchmark for future works.