||Nowadays, the Internet connects hospitals, medical institutions and people together. In this thesis, we focus on a medical application on ultrasound image processing and transmission. We consider that a pregnant woman takes her ultrasound examination in a local hospital, and the diagnosis and analysis of her ultrasound video (or image sequences) could be done remotely via the Internet. It is also possible for medical professionals and her family members from a distance to simultaneously view her ultrasound video using a multicast-capable network. In order to facilitate clinicians' diagnosis and improve medical information distribution, we address the issues in the clinical system related to the following two components. The first component focuses on automatic detection of fetal head features in ultrasound images. In clinics, fetal biometric measurement is done manually in the qualified ultrasound scanning images, in which some important predefined features exist. We present simple and new methods for automatic detection of these features in a single head ultrasound image. We also describe our system and criteria for ranking multiple images to facilitate clinicians' subsequent image selection and measurements in a sequence of ultrasound images. The results show that the methods can successfully identify features, and our system's ranking and the clinicians' ranking are well matched (with the hit ratio of 72.5%). The second component in the system is to offer quality real-time medical video to users over multicast networks. Since the video packet loss is inevitable, they have to be mostly recovered at clients. We propose and study an effective feedback-free recovery scheme for layered video in which the server multicasts FEC packets and replicated delayed (ReD) version of the streams in parallel with the video layers and the receivers, depending on their local losses, autonomously and dynamically join these layers to minimize their residual error rates (i.e., error after correction). Via analysis, we study the number of replicated streams and FEC packets that the server should provide and the optimal selection of FEC and ReD packets that a receiver should join. We further present fast approximation algorithms for the receiver to allocate the layer bandwidth, FEC and ReD packets, which are useful to those clients of low processing capability. We finally show that combining FEC with merely one or two stream replications can achieve better video quality on subjective experiments and substantially reduce the residual error rate as compared with pure FEC or replication alone (by as much as 50%).