Icometrix helps to detect and treat neurological diseases with AI imaging, powered by AWS

One in three people globally is thought to suffer from a neurological disorder at some point. In 2017, one study estimated that the cost of care for patients suffering from the nine most common neurological conditions amounted to nearly $800 billion in the US alone. For example, Alzheimer’s disease, the most common cause of dementia, affects six million people in the US, a number set to rise to 13 million by 2050. Multiple sclerosis (MS), another devastating neurological condition, affects an estimated three million people globally. These numbers illustrate the burden neurological diseases impose on society.

Making the correct diagnosis, not to mention prognosis, can be challenging due to the brain’s complexity. Icometrix, a healthcare technology company (HealthTech), aims to support neurologists in making the correct diagnosis and prognosis for several brain diseases. More than 500 clinical practices employ icometrix’s artificial intelligence (AI) imaging solutions to aid in detecting and treating MS and Alzheimer’s. To power their technology, icometrix uses Amazon Web Services (AWS).

Machine learning – focus on image segmentation accuracy

Both MS and Alzheimer’s feature significant brain volume loss (atrophy), with MS additionally marked by sporadic inflammatory brain lesions. To make an accurate prognosis and treatment decision, clinicians visually assess, rate, and count the brain volume loss and brain lesion load. Human variability and compounding pathologies make this process very error prone. To address this, icometrix developed a machine learning (ML)-based analysis pipeline. This pipeline allows radiologists to shift from subjective assessment of the MRI scans towards a quantitative, data-driven approach.

By 2015, however, icometrix noticed that these ML-based pipelines were reaching their limits regarding accuracy of the measurements and computation time. The latter is important for returning results in time. A new pipeline was developed using a hybrid approach. This new pipeline was based on a traditional unsupervised ML technique and a deep learning (DL) attention 3D U-Net. The attention 3D U-Net was specifically trained to address the weaker points of the traditional ML approach, namely difficulties in segmenting lesions in two different brain regions, which are important for assessing MS.

Deep learning – focus on image segmentation speed

The DL component of this initial DL and ML hybrid approach was built and trained in-house at icometrix using an attention gate 3D U-net, which guides the training process so that the network focuses on learning more salient features of the input. Because the algorithm is still hybrid, processing times remain significant.

The latest DL-based pipeline has reinvigorated the goal for more accurate and faster analysis of MRI scans. This new model decreased the measurement error of brain atrophy and white matter lesion detection by more than 30 percent. One of the most impactful changes is the reduction of processing time by 40 percent, resulting in much faster turnaround. This means that patients and neurologists have faster access to the results, reducing the time a patient spends waiting between the MRI being taken and seeing their neurologist.

Figure 1. Graphical representation of the attention-gate U-net used for segmentation and classification of brain lesions.

Due to the highly parallelized way in which typical DL models operate, this new DL pipeline allows the use of all available computational resources on a given instance. The older ML pipelines suffered from bottlenecks and did not scale very well. The models run on a CPU server without compromising computational speed, reducing the costs associated with each analysis.

AWS architecture of the icometrix DL inference solution

Compute

Icometrix recognized the potential of containers early on, giving them a competitive edge. However, the fragmented landscape of orchestrators posed challenges because Kubernetes had recently been released. Initially, icometrix used Apache Mesos to manage containers, but maintaining an orchestrator created a significant workload for icometrix’s IT department, especially as a startup.

Over time, orchestrators consolidated, leading to the release of Amazon Elastic Kubernetes Service (Amazon EKS). Icometrix migrated from Apache Mesos to Amazon EKS, reducing the IT team’s workload and allowing them to focus on the company’s core business.

Containerizing image analysis pipelines enhanced capabilities to meet growing customer demands. Noncontainerized pipelines would require specific coding for each computing environment, limiting scalability. In contrast, containerized pipelines, combined with Amazon EKS and Amazon EC2 Auto Scaling, enabled cost-effective scaling across various instance types. Utilizing this setup, icometrix can run a large number of distributed inference jobs with Auto Scaling groups, combining Spot, On-Demand, and Reserved Instances for cost efficiency and availability, which are critical in healthcare settings.

Figure 2. Architectural diagram of icometrix’s DL inference solution on AWS. The cloud solution is connected with on-premises hospital systems.

Storage

Patients of chronic neurological diseases often undergo longitudinal MRI analysis to track disease progression and treatment efficacy over time. However, managing and storing the resulting voluminous data can be challenging. This is where Amazon Simple Storage Service (Amazon S3) storage archive and tiering come into play, offering a comprehensive solution.

Amazon S3 provides a scalable and cost-effective storage solution for the volumes of longitudinal MRI data that are generated. By taking advantage of Amazon S3’s storage classes combined with the known scan intervals of chronic neurological diseases, icometrix can quickly move its data into optimized storage tiers.

Figure 3: Transitioning objects to more cost-effective tiers with Amazon S3 Lifecycle.

This storage setup works well in clinical use, but when training ML models, the access needs and patterns vastly differ.

For the DL training of the image analysis models (not pictured in Figure 2), icometrix uses Amazon Elastic File System (Amazon EFS). Amazon EFS automatically grows and shrinks file storage as organizations add and remove files without needing management or provisioning. This eliminates the requirement of manually allocating additional storage capacity, which is common with traditional storage mechanisms. Amazon EFS is particularly valuable during model training as it allows datasets to be expanded and shared simply between research instances, facilitating collaboration, performance, and scalability.

Conclusion

Continuous advancements in AI have significantly enhanced the accuracy of icometrix’s MRI segmentation models. Combined with AWS services, icometrix has achieved more accurate results and faster turnaround times.

For more information, visit icometrix.com or email their team.

To learn more about how to set up ML pipelines on AWS for AI-based medical imaging research, we recommend the following resources:

To learn more about how other imaging HealthTechs are using AWS, check out the following resources:

AWS Public Sector Blog