Leveraging Virtual Containers for High-Powered, Collaborative AI Research in Radiology

Abstract

Numerous obstacles confront radiologists interested in the use of artificial intelligence (AI) models within the field of radiology. For example, discrepancies between the radiologist’s and an AI developer’s hardware and software specifications pose a substantial hindrance to using AI models. Additionally, accessing and using GPU computers can lead to compatibility issues and add to these challenges. Finally, the dissemination of AI models and the ability to download pre-existing AI models are not simple tasks due to the size and complexity of most programs.

Virtual containers offer a solution to such compatibility issues and provide a simplified way for radiologists to use AI models. Virtual containers are software tools that bundle code, required programs, and necessary software packages to ensure that a program runs identically for all users, regardless of their computing environment. This article outlines the features of virtual containers (compatibility, versatility, and portability) and highlights an applied use-case for virtual containers in the development of an AI model.

Introduction

Artificial intelligence (AI)-powered tools and models are becoming increasingly important for the future of radiology [1]–[3]. Radiologists themselves are increasingly getting involved in evaluating the performance of AI models, deploying models, and even collaborating in the development of such models. However, the standard process for developing these AI models is fraught with challenges surrounding software incompatibilities and strict requirements for the coding environment. Many AI-based tools pull information and functionality from a variety of different base programs, and thus after a developer designs an AI tool, it is unlikely that a radiologist could seamlessly run the application. Additionally, accessing and using GPU servers to accelerate AI model development, as well as sharing AI models with other researchers, are generally complex tasks requiring time-consuming troubleshooting. Software incompatibilities can be mitigated with Virtual Containers (VCs). VCs are software tools that contain the necessary code and instructions needed to build the virtual environment required for a given AI model to function optimally.

This article will explore three benefits of VCs: compatibility, versatility, and portability. These benefits facilitate AI program development in radiology. A use-case for VCs will be described to demonstrate how a VC aided a radiologist in the development of an AI model to segment the kidneys from non-contrast CT images.

Compatibility

As the involvement of radiologists in AI model deployment and development becomes more commonplace, it is vital that tools are available to facilitate the involvement of radiologists. Collaboration on AI programs often proves exceedingly challenging owing to disparities in software specifications between radiologists and their collaborators. VCs offer a solution by bundling all necessary programs and their versions with the code, which facilitates seamless sharing of an AI model among users. This process ensures consistent functionality of the AI model among collaborators (Figure 1).

Figure 1.

Users with varying operating systems, applications, software versions, and other software configurations cannot easily share and collaborate on the same code. Sending code from one user to another without ensuring that the environment is compatible will likely result in errors and inconsistent performance of the AI model.

Versatility

Powerful computers, such as GPU servers, are a vital resource for AI researchers. However, these servers often lack pre-installed programs, such as Python, Java, or MATLAB. These machines require users to explicitly designate the software required for running the AI model. VCs automate this setup requirement by encapsulating the program’s environmental dependencies (Figure 2). Therefore, VCs simplify the software setup on servers for radiologists and their collaborators.

Figure 2.

GPU servers do not typically have programs and software packages installed. Thus, the code environment from a user’s computer cannot seamlessly run on a GPU server unless such programs and software packages are installed when the code is sent to the GPU server.

Portability

The development of an AI model de novo is a time consuming and challenging process and therefore typically lies beyond the bandwidth and/or expertise of most radiologists. Yet, AI specialists in recent years have developed a plethora of AI models, which are often publicly available. Radiologists can retrain these available models using new data and repurpose the AI model to address unique inquiries. Virtual containers, such as Docker (Docker, Inc.; Palo Alto, CA, USA), facilitate the seamless sharing and adaptation of AI models through cloud-based repositories. These virtual containers lower the barrier to entry for radiologists to use AI models. Radiologist researchers can thereby capitalize on this existing body of work and expand upon proven frameworks to suit their specific needs.

Once an AI model is trained, the code can be uploaded to a VC repository and made publicly accessible (Figure 3). Subsequently, radiologist researchers can download the VC from the repository and execute the code exactly like the original developers, notwithstanding any specific hardware requirements such as GPU availability or storage capacity (VCs cannot replicate GPU functionality or virtually expand storage capacity).

Figure 3.

Virtual containers allow code to be packaged for distribution. Information regarding the operating system, programs, and software variables are seamlessly integrated into the container. The code within the container will operate smoothly for other users irrespective of the software on the recipient’s computer.

Virtual Container Options and Alternatives

Several implementations of VCs exist, including Docker, Kubernetes, Podman, and Singularity. Docker has emerged as a leading platform for creating containers and allows users to incorporate all dependencies and libraries in the code [4]. Kubernetes, often paired with Docker, excels as an open-source platform for automating the deployment, scaling, and management of application containers. Kubernetes excels at container deployment and operation [5]. Podman, which offers functionality akin to Docker, has an architecture beneficial in certain computing environments, which is outside the scope of this article. Finally, Singularity is a container primarily employed in scientific and high-performance computing domains that has less stringent security requirements [6]. Each platform presents distinct features suitable for various applications, with Docker and Kubernetes standing out due to their extensive feature sets and robust community support.

Virtual Machines (VMs) are an alternative means of running software applications. VMs are instances of an operating system that run software, whereas VCs are a compartmentalized application on the same operating system. VCs have several advantages over VMs, and specifically are more portable, quicker to start up and shut down, and utilize fewer resources.

Use-case: Virtual Containers for Implementing a Deep Learning Model for Segmentation of the Kidneys on Non-Contrast CT Images

CT image segmentation finds applications in the identification and interpretation of pathology as well as large-scale research projects. Numerous publicly available deep learning models offer highly accurate organ segmentations [4]. In this study, a segmentation AI model was acquired and adapted to operate on a GPU server using a VC. The selected segmentation model was originally designed to identify the kidneys on contrast-enhanced CT images. The radiologist collaborator desired to implement this AI model and retrain the segmentation model to perform on 210 non-contrast images.

For this use-case, Docker software (version 24.0.5; Docker, Inc.; Palo Alto, CA, USA) was chosen as the VC platform [7]. We used a departmental GPU server, instead of a local/personal GPU, which provided significantly greater computing power. However, akin to many GPU servers, this server lacked the necessary programs required for proper code execution, such as Python (version 3.8; Python Software Foundation; Delaware, USA).

A folder was then made, containing the three essential elements required for deploying a virtual container: the code, the data, and the setup instructions. The code selected was an AI-driven segmentation algorithm (specifically the MIScnn package from the KiTS19 challenge (1,2)) sourced from the Docker repository, which eliminated the need for creating a new AI model from scratch. The dataset consisted of non-contrast CT scans of the abdomen and pelvis. The setup instructions detailed the AI model’s environment by specifying the operating system (Linux), programming language (Python), and necessary libraries (NumPy, TensorFlow, etc.).

The provided instructions were critical for the functionality of the virtual container. The instructions ensured precise execution of the program, even in the absence of preinstalled software on the GPU server. Despite the server lacking the necessary programs, the VC was the sole prerequisite for successful operation.

After assembling the folder with the code, data, and environment instructions, we initiated the Docker program on the GPU server. Within approximately 12 hours, the program completed its execution. If performed manually, each segmentation would have required at least 30 minutes to import, segment, convert between file types, and export. To do this task consistently for 210 studies would therefore take around 100 hours. Subsequently, we exported the training results and the optimal segmentation model from the Docker container to the radiologist’s local computer for further analysis.

Discussion

Although the implementation of an AI model is possible without using virtual containers, the deployment of the model would have been otherwise difficult and likely prohibitively time consuming for the typical radiologist. Virtual containers also aid in collaborating on code development, as the container mitigates conflicts stemming from disparate information on individual users’ local machines. As a result, virtual containers effectively guarantee that the execution of code is standardized across diverse user environments.

The segmentation AI model deployed in this example was derived from code accessed through a virtual container code library. This process thereby represents a typical scenario for the radiologist, who desires to use a publicly available model on either local data and/or for a new application. The use of a comprehensive code library and associated virtual containers allowed the radiologist to use these models with relative ease. Otherwise, the development of these models would pose considerable challenges and potentially need the expertise of specialized AI computer scientists.

In the use-case, VCs demonstrated three pivotal advantages. Firstly, VCs provided compatibility, which ensured efficient collaboration among lab members and mitigated errors in software compatibility (Figure 4A). Secondly, the versatility of VCs allowed us to leverage a GPU server, which enhanced computational power and expedited data processing for the use-case (Figure 4B). Lastly, the portability of VCs facilitated the use of a pre-existing AI model via the VC repository, which allowed for the customization of the AI model to the use-case. In turn, the re-implemented AI model can be easily shared and utilized by others using a VC (Figure 4C).

Figure 4.

A. Compatibility: Two researchers can use virtual containers to collaborate on a project, irrespective of the coding environments on their computers. B. Versatility: A researcher can use virtual containers to access remote GPU servers to process data more efficiently. C. Portability: Once a program is published as a virtual container, other users can remotely access the program and recreate the operating conditions of the original authors.

Conclusion

Virtual containers (VCs) offer invaluable support to radiologists and radiology researchers in the development and implementation of AI models, as VCs provide three core features: (1) Compatibility, (2) Versatility, and (3) Portability.