HEALTHCARE OUTCOMES

While studying advanced 3D imaging and AI in healthcare applications at Taiwan’s National Tsing Hua University, researchers hit on an exciting potential application: helping pathologists diagnose cancer tumors with greater speed and precision. They obtained licenses from the university and formed a digital imaging startup, JelloX Biotech Inc., but soon discovered hospitals were far from ready to adopt the technology.

In the age of precision medicine for cancer treatment, growth rate of well-trained pathologists is far beyond the need of diagnosis. Most pathologists still examine tissue samples by eye and take manual notes, a painstaking hours-long process. Few have made the switch to digital 2D or 3D image analysis, in part because it traditionally has required installation of costly and complicated graphics equipment.

Computer Vision in Healthcare

Despite their highly trained eyes, doctors don’t always get important details right. Tumor samples are complex—each one contains 10 to 30 parameters that must be analyzed to determine whether the cells are cancerous, how fast they are dividing, and how healthy or unhealthy they look as compared with normal tissue, among other factors.

“Studies asking multiple pathologists to analyze the same tissue sample have found 20% to 30% disagreement among the diagnoses,” says Yen-Yin Lin, Chief Executive Officer at JelloX. “This means that there is a chance that patients might receive incomplete information about their disease status, thus delaying proper treatment.”

Misdiagnosis can be very painful for patients. They might miss a good chance to use the best drug for fighting their cancer earlier or undergo chemotherapy they may not need.

To improve diagnostic capabilities without breaking the bank, Lin and his colleagues set out to create an edge solution that could quickly uncover and digest far more information than pathologists can see—without the need for installing expensive graphics equipment.

“#AI insights could help doctors improve diagnostic accuracy and develop better #treatments.” – Yen-Yin Lin, JelloX Biotech Inc. via @insightdottech

Using AI 3D Imagery in Pathology

The company found recent advancements in computer vision and AI could be used to help doctors better detect anomalies from medical images with higher accuracy. “It can assist healthcare professionals in diagnosing diseases like cancer, identifying disease progression, and predicting patient outcomes,” says Lin.

As a result, JelloX set out on a three-year development journey to create MetaLite Digital Pathology Edge Solution, which can analyze each tissue sample parameter in one to two minutes, compared with an hour using a standard computer.

To do this, JelloX needed to leverage powerful deep-learning models and annotation tools, which required equally as powerful hardware capable of deploying these models at the edge and reducing inferencing time for quick, efficient, and accurate results.

Lin explains they turned to an edge computing device powered by Intel^® processors and custom AI algorithms deployed through the Intel^® Distribution of OpenVINO^™ Toolkit. This made it highly suitable for deployment on netbooks.

Intel CPUs were able to accelerate the training and inferencing significantly, provide an end-to-end deep-learning pipeline that helped JelloX apply its solutions to real use cases, and deploy their models across different hardware.

This is because OpenVINO was designed to first help to optimize deep-learning models, then deploy the model over multiple hardware devices, and accelerate the inference and performance of those platforms, Zhuo Wu, a Software Architect at Intel who works closely on OpenVINO, explained.

As a result, JelloX can now help configure most hospital scanners to work with the software, which also allows doctors to add notes as they work (Video 1).

In addition to Intel CPUs, JelloX also leverages Intel^® NUC based on the 11^th Gen Intel^® Core^™ processors, which enable engineers to easily scale their solutions.

Pathologists can choose to review some parameters in real time and save others for later. Data from the scanner and edge device is sent to hospital servers, where hundreds of parameters can be analyzed with AI in detail overnight, with results ready to view the next morning.

AI models are trained on massive data sets accumulated from many sources. The amount of information they work with is too vast for humans to assimilate, but algorithms can quickly process it and use it to classify tissue samples and make inferences and predictions about the course of the disease.

“The interpretation of immunohistochemistry staining is a time-consuming and expensive process in pathological examinations, requiring significant time from physicians. If auxiliary tools can be utilized to improve efficiency, it can bring about the greatest economic benefits. Some parameters are difficult for doctors to categorize conclusively. When AI does calculations, it gives doctors a scale or digital ruler to use as they judge the images.”

AI insights could also help doctors improve diagnostic accuracy and develop better treatments, Lin believes, saying, “If we have good AI-assisted tools, maybe patient survival rates and survival duration will be enhanced.”

AI analysis is also valuable to medical researchers, allowing them to discover new features of cancer cells and better understand how they operate. “Algorithms can dig out more information from images and do the tough analysis, providing more information about morphology and protein biomarker features,” Lin says.

Being able to gain more efficient and accurate results with AI not only helps doctors improve patient care and service but also reduces the time and effort they need to spend on each case—which in turn allows them to take better care of more people, according to Wu.

Currently, researchers at Taipei Veterans General Hospital and MacKay Memorial Hospital in Taiwan are using MetaLite to identify new biomarkers of cancerous tissue and calculate the area of tumors with greater precision. Once the platform receives approval from Taiwanese health authorities, the hospitals may use it as a diagnostic tool.

Pharmaceutical companies may also benefit from AI tissue analysis, using it to identify which patients stand the best chance of benefitting from medications set to undergo clinical trials, especially for those requiring biomarker-guided patient screening.

Expanding AI in Healthcare with Federated Learning

As hospitals expand the use of AI in pathology, data they obtain will be used to train future AI models, increasing accuracy. And through a process known as federated learning, hospitals can now securely share image data with others while confining sensitive patient information to their own servers—a capability once considered an impossible dream. JelloX is developing a new version of its software that enables federation.

“With federated learning, data will accumulate much faster, improving the AI and increasing speed and data uniformity,” Lin says. “Using AI in pathology will drive precision medicine, helping doctors improve diagnosis and treatment, and allowing pharmaceutical companies to develop new drugs much faster.”

In fact, in its immunohistochemistry imaging solution, the company is already leveraging Intel’s open-source framework Open Federated Learning (OpenFL) to enable seamless cross-institutional analysis of images with many of its customers.

“AI is becoming more prevalent in the healthcare space due to its immense potential to revolutionize healthcare delivery, improve patient outcomes, and enhance operational efficiency,” Lin adds.

Beyond federated learning, AI is also coming to healthcare in the form of chatbots and virtual assistants—enhancing patient engagement and support. Using natural language processing, conversational AI chatbots can help collect accurate patient information so doctors and nurses can focus better on patient care, according to Wu.

To learn more about developing healthcare AI solutions, check out these notebooks: Quantize a Segmentation Model and Show Live Inference and Part Segmentation of 3D Point Clouds with OpenVINO^™.
 

This article was edited by Christina Cardoza, Associate Editorial Director for insight.tech.

This article was originally published on September 22, 2022