Imaging Diagnostics for Population Screening — Role of Edge Devices and Cloud Machine Learning
Edge devices are revolutionizing diagnostics. An edge device is a system that allows computing to happen closer to the location where it is needed, resulting in an instant response. Edge devices can be attached to imaging tools such as digital X-ray, CT, MRI, or ultrasound equipment. They can also be deployed as screening point-of-care (PoC) solutions. Some advantages of edge devices include instant results, no latency, and remote deployment for pre-screening. However, they suffer from a low computation capacity, and the number of algorithms that can be deployed on them is limited.
In population health, tuberculosis screening provides an example where digital chest X-rays play a pivotal role in disease assessment and management. Without the use of technologies like AI and edge devices, the medical images obtained are either pushed by encrypted protocol to an expert remotely or physically reported by taking hard copy prints of images. In such a system, a study cannot be prioritized unless the imaging expert either reads and reports all studies in the order in which they were obtained or manually searches for abnormal studies. This Workflow management is non-optimized and leads to reporting delays, stress, and burnouts.
Point-of-care solutions that allow testing and pre-screening near the patient mitigate these issues. The AI models prioritize studies instantaneously so that patients can be held back for additional investigations. This reduces the number of patients lost to follow-up and the cost of follow-up while controlling the spread of infection to other people in society. The overall burden of stress on imaging experts and other healthcare workers is also significantly reduced. Porting these models on edge devices allows us to deploy AI solutions remotely. This is a game changer in population health screening.
On the flip side, a cloud-deployed algorithm has distinct advantages. The AI algorithm residing in a cloud model uses high computing power to run AI-based triage and other functions for multiple conditions. This may not be possible on the nano data center on an edge device due to its limited computing capacity.
Edge device learning coupled with cloud machine learning is a unique model. Edge devices can give an instant diagnosis and run processes that require low computation power. Cloud-based tools can use their high-capacity computing power to use additional algorithms on the same X-ray image and run image analytics. A system’s reliability becomes hard to question when cloud services are used in conjunction with edge devices to eliminate break downs.
The Greater Chennai Corporation’s active screening for TB under the TB Free Chennai Project, using mobile diagnostic vans (MDUs) and DeepTek’s Genki solution, is an example of deployment of cloud and edge solutions to pre-screen and triage for TB. The AI triage results were reviewed by an imaging expert in the loop and further validated. The model assessed and classified 75,000 X-ray images into two classes accurately: (1) patient likely to have TB and (2) patient unlikely to have TB. All these images were later evaluated by an expert radiologist.
A hybrid model combining edge, cloud, and teleradiology solutions, is yet another enhanced model. It allows an imaging expert in the loop to edit and update reports and re-categorize studies. This feedback helps the model learn continuously and better itself.
Edge diagnostics and edge learning devices are going to power the devices of the future in the medical imaging space. In their current form, these devices perform basic functions such as triage and pre-screen studies. As high computing power makes deep learning solution deployment easier on edge devices, multi-class and multi-function models are likely to be deployed on edge devices. The democratization of radiology could also be a reality if edge devices are deployed to perform basic triage functions in remote areas where medical imaging specialists are unavailable.
Read the full paper at https://arxiv.org/abs/2006.13808.
