Exploring the Potential of AI to Advance Diagnostic Pathology

Feature

Exploring the Potential of AI to Advance Diagnostic Pathology

Body

Kenji Ikemura was born with an analytical mind and a compassionate heart. As a teenager growing up in Tokyo, Japan, he decided that he wanted to become a doctor. But he also had a passion for engineering and basic science. So, when it came time for college, he moved to Irvine, California, and earned a degree in biomedical engineering from UC Irvine. A few years later, he received an M.D. degree from Rush Medical College in Chicago.

“I knew that I could do my best work as an engineer in a medical environment,” explains Dr. Ikemura, who is now a second-year Pathology resident at Montefiore Einstein. “I wanted to make medicine more efficient for patients.”

Dr. Ikemura’s desire wasn’t just theoretical, it was personal: A family friend back home in Irvine had been suffering from breast cancer for years and her condition was getting worse. “I wanted to do something that would marry my interests in Pathology and engineering and somehow help people like my friend Beth,” he says.

An Aha! Moment

During his fourth year of medical school, Kenji had an idea: What if a physician could use their cell phone to make a breast cancer diagnosis?

He set about devising a way to merge Artificial Intelligence (AI) with pathology, with a specific focus on breast cancer. The first step: to train object recognition software to recognize cancer features on histology and detect mitoses. What began as a self-described “hobby” he worked on during his “free time” in medical school has evolved into a research project aimed at developing a mobile phone app that can be used to classify breast cancer histology images in real time.

He named his project, The Beth Project.

Finding Collaborators

After arriving at Montefiore in 2018 to begin his residency in Pathology, Dr. Ikemura approached faculty member Susan Fineberg, MD, an anatomic pathologist who specializes in breast cancer. He needed to know if the data he’d initially collected from online public resources were accurate.

Dr. Fineberg was intrigued. She enlisted the help of Dr. Miglena Komforti, a pathology fellow specializing in women’s health. Drs. Fineberg and Komforti confirmed that the images Dr. Ikemura had collected were accurate and could be used to train the AI algorithm. Then they showed him how to classify the different types of breast slide images: benign versus carcinoma in situ, or invasive carcinoma.

Problem Solver on a Mission

Deep Learning (DL) is a branch of AI that models the human brain’s neural network. You can train the computer to do a specific task without hard-coding the instructions, Dr. Ikemura explains. Convolutional neural network (CNN) is a type of DL optimized for image recognition.

Dr. Ikemura aims to determine whether CNN can be applied for histological image analysis, and if its network can be transferred to a commercially available mobile phone for real-time histological image analysis captured through the cell phone camera.

“If AI is in your pocket and you can hook it up on a microscope, then you don’t need an expensive slide scanner. You just need your phone,” Dr. Ikemura explains. The app can be used as a decision-supporting system for the pathologist, he says, and a cell phone can also connect to the internet, thereby making telepathology possible. This could be especially helpful in caring for people in rural areas where pathologists are not readily available.

Breast cancer histology image from mobile-phone camera. AI output states with 82.6% confidence that it is looking at invasive breast cancer.
Breast cancer histology image from mobile-phone camera. AI output states with 82.6% confidence that it is looking at invasive breast cancer.

Next Steps

Making a diagnosis often requires various steps, such as counting mitoses. This can be a tedious process that involves moving the microscope from one field to another, multiple times. “If you could hook up a phone and have it detect and count all the mitoses,” says Dr. Ikemura, “it would make the pathologist’s job easier and possibly more precise.”

But for now, he wants to make sure the breast pathology works. He’s hoping to plot more data, refine the AI algorithm, and determine not only whether it can detect the type of tumor, but that it can also identify and count the mitotic figures within the tumor. Dr. Fineberg notes that “the technology Dr. Ikemura is developing, which would allow for more accurate mitotic counting, could have a tremendous impact on patient care.”

Bringing AI to the Hematology Lab

Digital imaging has revolutionized the pathologist’s job of looking at cell specimens through a microscope to render a diagnosis. During his rotation in hematology as a first-year resident, Dr. Ikemura learned from Morayma Reyes Gil, MD, PhD, director, Montefiore Einstein Clinical Hematology and Coagulation Laboratories, that the lab was having challenges with CellaVision, its digital imaging system. Dr. Ikemura offered to help.

Ideally, the CellaVision software takes photos of patient blood specimens and pre-classifies white blood cells according to the different types (monocytes, neutrophils, lymphocytes, eosinophils and basophils). But Dr. Reyes Gil noticed a glitch: While CellaVision is able to differentiate among normal blood cells, she says, it does a poor job of identifying abnormal cells such as leukemia cells (also known as blasts or immature blood cells) and classified them as “unknown” or  “other” cells and sometimes as “junk”.

When presented with this image from CellaVision, the AI is 99.8% confident that it is NETS.
When presented with this image from CellaVision, the AI is 99.8% confident that it is NETS.

“This is a serious problem,” says Dr. Reyes Gil. “Clinicians depend on us to identify cases of life-threatening blood disorders such as leukemia. When a patient presents with fevers or weight loss, we’re the first to see the malignant cells in the blood.” Speed and accuracy in rendering a diagnosis are critical.

In discussing the problem with Dr. Ikemura, Dr. Reyes Gil posed two questions: Could AI be used to enhance CellaVision’s ability to classify different types of white blood cells? Could the AI technology Dr. Ikemura developed for breast cancer histology be trained to make the leap from solid tissue to blood? Together, they stepped up to the challenge.

Novel Research

As they began their investigation, Drs. Ikemura and Reyes Gil found neutrophils—cells on the frontline of fighting infection—among the cells that CellaVision discarded as “junk”. When neutrophils eject their nuclear material (chromatin) through a process called “NETosis,” bacteria-killing “neutrophil extracellular traps” (NETs) are formed.

Immunologists and basic biologists have known about NETosis for over 20 years, says Dr. Reyes Gil, but the main studies have been done in animals and in tissues. Before she and Dr. Ikemura began their project, she notes, no one had looked into identifying these cells in blood smears.

Dr. Reyes Gil established biomarkers to associate certain conditions with high levels of NET activity in patient blood samples. Third-year pathology resident Dr. Mohammad Barouqa assisted with the characterization of the NETs in blood, and with staining and developing protocols to use as a positive control. Dr. Ikemura modified his software and taught it to identify NETs.

In examining cells that CellaVision was discarding as junk, the researchers observed some that appeared to have a high degree of NET activity. These cells tested positive for the biomarkers, confirming the likelihood they were NETs. They took photos of the cells they knew were not NETs, and of those that they suspected were NETs. The AI tested 98 percent for accuracy.

Confident they were on the right track, Drs. Ikemura and Reyes Gil invited Henny Billett, MD, chief of the Montefiore Einstein Department of Medicine’s adult hematology division, to collaborate with them on a blind study. Dr. Reyes-Gil grouped cases she saw randomly coming from CellaVision with a high number of NETs, versus cases with no or low numbers. Dr. Billet and her team assessed the data to determine whether the patient had an infection (bacterial or viral), a recent malignancy, renal condition, liver disease or immune disorder. There was a significant correlation between the high-NET cases and bacterial and viral infection.

Going Public

The clinical study was the first of its kind. Dr. Billet presented a poster at the American Society of Hematology meeting in December 2019. Dr. Reyes Gil gave an oral presentation at the International Medical Pathology meeting in February 2020 and has submitted two abstracts to the International Society of Thrombosis and Haemostasis. Drs. Ikemura and Barouqa will present at the Academy of Clinical Laboratory Physicians and Scientists meeting in May.

With the assistance of Einstein’s Biotechnology and Business Development team, Drs. Reyes Gil, Ikemura, and Barouqa applied for and received a provisional patent from the United States Patent and Trademark Office in July 2019. They have one year to produce additional data to make the project eligible for a full (national and international) patent.

Q & A with Dr. Kenji Ikemura and Dr. Morayma Reyes Gil

Five Questions for Dr. Ikemura

Why did you choose to specialize in Pathology?

To me pathology is the most fundamental truth in medicine. It’s the foundation of medicine. It’s like mathematics to engineering. As a pathologist, you actually hold the specimen in your hand and are able to analyze it. That really made me interested, not only as a diagnostician but as a clinician who will be treating patients.

It all burns down to what we see under the microscope and what we analyze at the cellular and DNA level. When we know and understand the pathology, the mechanism of disease, we can treat the patient in a way that is specific to that patient.

Personalized medicine is the wave of the future. I believe that as personalized medicine evolves, Pathology will play a greater role in developing more-effective treatments against cancer.

Do you think that AI could eventually make the role of the pathologist obsolete?

As we continue to train the AI, it will get better and better at what it does. AI won’t make a final diagnosis and won’t replace the pathologist. Rather, it can be a tool to help guide the pathologist in rendering an accurate diagnosis.

How do you envision your future medical career?

I would like it to include some level of clinical work, to help me understand the problems patients are facing. Once I understand where the inefficiencies lie, I can apply my engineering skills to make the clinical care more efficient.

Pathology residency is very demanding. What keeps you motivated to continue working on the Beth Project?

Beth died in December 2018. Before she passed, we talked about the Beth Project. She was proud of the work. One in eight women gets breast cancer. That infuriates me. As I work on this project, Beth is always on my mind. She is my motivation. I can’t stand the thought of someone as kind and as beautiful being in agony.

What is novel about the Beth Project?

Researchers around the world are exploring the application of AI in medicine, in general, as well as in pathology. What I’m trying to do is move that AI into the mobile phone, so that AI with the ability to diagnose cancer is in our pocket. From my literature search, I haven’t seen a study like this yet, so that is the novel aspect of my project.

Five Questions for Dr. Reyes Gil

What is the relationship between medicine and AI?

AI is the future of medicine, there is no question about it. The question is, why is the use of AI in medicine so far behind?  After all, AI has been well adapted in other fields. I think part of the reason is the resistance from physicians to accept that something artificial like a computer can have better judgment than a doctor, and the fear that we can be replaced. It’s a naïve concept. Those who have the fear are not exposed to how AI really works.

It’s impossible to think that a single doctor can know all the answers. It’s old-fashioned thinking. We need to use algorithms that will allow us to m ake better medical decisions based on the information available. There’s just too much information for our brains to process. AI can help us get the most pertinent information faster and save lives.

What is the relationship between Pathology and AI?

AI is the future and we need to embrace it. Given that Pathology is at the forefront of medical informatics, not only should we embrace AI, we should be the leaders in adapting it to the medical field.

AI is a tool. We pathologists have to decide how to use this tool. I see us as the future engineers in developing algorithms in medical informatics to guide medical decisions utilizing all clinical and laboratory data. We’re the gatekeepers of so much medical data, we need to design better algorithms and operations to make it available in an organized and expedited way to facilitate medical decisions.

How does an engineering background like Kenji’s mesh with the field of Pathology?

These days It’s not too unusual to see medical students with backgrounds in engineering or computer science. Pathology allows them to continue to exploring the application of their knowledge. Computer science, particularly informatics, is essential for Pathology operations. So it’s a natural path.

Does the Montefiore Einstein Pathology Department offer learning opportunities for other residents interested in AI?

Dr. Evan Cadoff directs the informatics course in Pathology and Dr. Amy Fox directs the data analysis projects that all residents do as part of their training. Kenji and I have discussed the possibility of starting a journal club. Our Pathology residents have been trained to use data analysis and algorithms to make clinical decisions, and our faculty plan to expand on this, ideally they will have the opportunity to use AI in future projects.

Do you and Dr. Ikemura have any more AI projects in the pipeline?

We’re working on a project to train the AI to differentiate blasts from acute lymphoblastic leukemia (ALL) from that of acute myeloid leukemia (AML).