Artificial intelligence (AI) has been deeply woven into modern-day healthcare ranging from disease visualization to aiding medical decision making. However, the use of AI in medical imaging comes with certain challenges. In this post, we look at one of the pivotal challenges – data privacy – and examine a framework we designed that addresses this concern while deploying deep learning algorithms using edge computing.

The Need for a Solution

Data privacy has transformed into one of the major concerns when employing deep learning systems in clinical health practice, especially through cloud computing. It’s vital to maintain a balance between high flexibility (like cloud computing) and security (like local deployment) without risking the exposure of Patient Health Information (PHI).

Current solutions offer a mix of confidentiality and convenience. Bespoke desktop software solutions demand a long, administrative approval process and are less scalable due to manual installations. Remote servers can be equipped with ample computing resources, but they necessitate the transfer of PHI from the clinic machine to the remote one, posing security risks. Finally, programs on the clinic machine possess neither of these disadvantages, but they often do not have access to scientific computing hardware such as GPUs.

Introducing Serverless Edge-Computing

For us, the answer lay in Serverless Edge-Computing. In contrast to server-based computing, where computation takes place on a central server, edge computing pushes the computation as close to the data’s source as possible. This allows heavy computations to be performed closer to the end device, reducing latency, and ensuring data privacy.

Our goal was an implementation that tackles the demanding task of 3D medical imaging by deploying a 3D medical image segmentation model for computed tomography (CT) based lung cancer screening.

Components and Functioning

Our implementation is a browser-based, cross-platform, and privacy preserved system. All computing operations, including data pre-processing, model inference, and post-processing, occur on user’s local devices without any data transmission or persistent data storage on the platform.

Here’s a quick look at the process:

• Pre-Processing: The 3D image volumes are loaded and converted to tensors. They are then scaled, reoriented, and padded.

• Model Inference: Once the tensor is prepared, it is fed into the model inference session.

• Post-Processing: The final phase involves storing the output back into a large volume tensor and removing padded voxels.

Performance Evaluation

We ran tests to characterize the runtime and memory usage of our solution on various devices with different operating systems including Linux, Windows, and macOS. Operating systems were tested on different browsers including Firefox, Chrome, Microsoft Edge, and Safari.

Our implementation achieved an average runtime of 80 seconds across Firefox, Chrome, and Microsoft Edge and 210 seconds on Safari. The average memory usage was also catered to a broad consumer base with an average use of 1.5 GB on Microsoft Windows laptops, Linux workstations, and Apple Mac laptops.

Current Limitations and Future Plans

Our design currently carries some limitations. Currently, many deep learning models require hardware acceleration or have memory usage that exceeds limitations imposed by web browsers. Our model inference runtime is also influenced by the number of threads, which is also another avenue for future system optimizations.

Despite these challenges, our framework effectively minimizes the risk of PHI exposure and demonstrates that a stateless, locally executed, and browser-based strategy is feasible and advantageous in the context of regulatory barriers and scalability.

Conclusion

The implementation of serverless edge-computing in AI-led medical imaging is a big leap towards a more secure and efficient healthcare ecosystem. As we continue to improve and develop the system, we are optimistic about the potential of these techniques to revolutionize medical imaging and bring greater value to healthcare providers and patients.

Tags: #ArtificialIntelligence, #MedicalImaging, #EdgeComputing, #DataPrivacy
Reference Link