There is an increasing interest among scientific communities for sharing data and applications in order to support research and foster collaborations. Interdisciplinary domains like neurosciences are particularly eager of solutions providing computing power to achieve large-scale experimentation. Despite all progresses made in this regard, several challenges related to interoperability, and scalability of Distributed Computing Infrastructures are not completely resolved though. They face permanent evolution of technologies, complexity associated to the adoption of production environments, and low reliability of these infrastructures at runtime. This work proposes the modeling and implementation of a service-oriented framework for the execution of scientific applications on Distributed Computing Infrastructures taking advantage of High Throughput Computing facilities. The model includes a specification for description of command-line applications; a bridge to merge service-oriented architectures with Global computing; and the efficient use of local resources and scaling. A reference implementation is proposed to demonstrate the feasibility of the approach. It shows its relevance in the context of two application-driven research projects executing large experiment campaign on distributed resources. The framework is an alternative to existing solutions that are often limited to execution consideration only, as it enables the management of legacy codes as services and takes into account their complete lifecycle. Furthermore, the service-oriented approach helps designing scientific workflows which are used as a flexible way of describing application composed with multiple services. The approach proposed is evaluated both qualitatively and quantitatively using concrete applications in the area of neuroimaging analysis. The qualitative experiments are based on the optimization of specificity and sensibility of the brain segmentation tools used in the analysis of Magnetic Resonance Images of patient affected by Multiple Sclerosis. On the other hand, quantitative experiments deal with speedup and latency measured during the execution of longitudinal brain atrophy detection in patients impaired by Alzheimer's disease.
