Applications in ATMOSPHERE are a combination of several levels of services involving the actual medical imaging applications and the execution environment. On the top, application
developers create medical-image processing applications for model creation, data analysis, and production. The requirements and architecture of each kind of application are quite different, requiring intensive computing resources in the data analysis and model creation phase. All these applications may require IPR protection and should not be available in public
repositories. Moreover, the application model is also different. Data analysis and model creation applications usually are single-tenant state-full interactive applications, and production applications are typically multi-tenant and stateless Web Services. As an example, an ATMOSPHERE application is a container running a classifier inside a Web Service deployed on an elastic Kubernetes cluster, but it is also a LEMONADE Machine Learning framework deployed on a Kubernetes cluster and linked to an elastic Mesos cluster running Spark jobs, all of them connected to a shared storage. Therefore, ATMOSPHERE will provide application templates and best practices to deploy them on the cloud services.

ATMOSPHERE identifies three types of architectures for medical imaging applications, two for the data analysis and model creation and one for the production. These applications combine different dependencies in software components with different restrictions and requirements. Applications may include functions for the provisioning of metrics to the trustworthiness system and the adaptation of the application. The development cycle of the application must be simple and systematic, to facilitate measuring the trustworthiness properties, reusing robust and secure components and automatic the deployment cycle. For this, ATMOSPHERE defines a Continuous Integration model with three levels: first, a set of trustful Docker images with the basic dependencies from different applications, without the specific proprietary code, automatically created, verified and uploaded in a public repository. Second, Docker containers with the application code created, tested and registered on a private repository. Finally, a set of TOSCA recipes that describe different application topologies are customized to deploy the schedulers and resource management services and to run the applications. The whole process aims at providing a convenient and effective procedure for supporting the development, testing, and deployment of applications in the ATMOSPHERE cloud.