Machine Learning Increases Diagnostic Accessibility, Saves Time in Histopathology

In Partnership With:

R. Lor Randall, MD, FACS, shares the successes achieved so far with machine learning in histopathology, explained the positive effect this technology can have on institutions, and advocated for further use of technology to aid treatment advances.

R. Lor Randall, MD, FACS

R. Lor Randall, MD, FACS

As artificial intelligence is advancing in depth and scope, histopathologists can use this technology to quicken the diagnostic process and expand access to pathology resources, according to R. Lor Randall, MD, FACS.

In a paper recently published online in Modern Pathology, Randall and coauthors demonstrated that digital systems could aid in recognizing histopathologies for a variety of rhabdomyosarcoma samples, supplementing work done by diagnostic specialists.1

“Machine learning and the greater scope of artificial intelligence is [putting us on] the cusp of being able to do pattern recognition at the cognitive level,” he said. “Histopathologists and all other diagnosticians are still absolutely critical to the mission of diagnosing and treating people with afflictions. [With these advances, we’re simply asking]: Can this technology aid us in the fidelity of the [histopathology] process and make this expertise available to people who may not have it locally?”

In an interview with OncLive®Randall, the David Linn Endowed Chair for Orthopaedic Surgery, a professor in the Department of Orthopaedic Surgery, and chair of the Department of Orthopaedic Surgery at University of California Davis Health, shared the successes achieved so far with machine learning in histopathology, explained the positive effect this technology can have on institutions, and advocated for further use of technology to aid treatment advances.

OncLive®: What was the rationale for the study you coauthored on machine learning in rhabdomyosarcoma histopathology and what were some of the most important findings?

Randall: This paper, engineered by Charles Keller, [MD, of Children’s Cancer Therapy Development Institute], looked at machine learning for rhabdomyosarcoma histopathology. The big pictures for this topic are that sarcomas are rare, there’s not a lot of sarcoma histopathology expertise around the United States or the world, and we live in an advancing digital era of artificial intelligence and machine learning.

Dr Keller, who is always on the forefront of thinking about sarcomas and pediatric oncology, gathered investigators from all over the world, obtained several pathology specimens, and, with some of his colleagues, developed a convolutional neural network-based differential diagnosis system of machine learning. This is basically advanced pattern recognition and machine learning such that the artificial intelligence becomes increasingly skilled with more exposure.

Looking at 424 specimens, the researchers trained this digital system to recognize histopathologies for alveolar rhabdomyosarcoma, embryonal rhabdomyosarcoma, and clear cell sarcoma. This research resulted in a receiver operating characteristic area under the curve value above 0.889 [for all sarcoma subtypes that were tested], which is a robust number.

They then took additional specimens, both from patient samples and genetically engineered rhabdomyosarcoma mice, and looked at the computer’s ability to recognize the histologies. They came up with some excellent numbers in the alveolar rhabdomyosarcoma samples, and not quite so good numbers with the embryonal rhabdomyosarcoma and clear cell sarcoma samples. The alveolar rhabdomyosarcoma histology was 0.89, [showing that the] computer was able to recognize histology with remarkable fidelity for these samples. The embryonal rhabdomyosarcoma histology was 0.61 and the clear cell sarcoma histology was 0.64.

There’s some room for improvement with some of the histologic subtypes. This is promising early work. We’re on the cusp of being able to use technology in diagnostics, in health care, and then ultimately in therapeutics through machine learning.

Can these diagnostic results improve as machine learning refines its ability to accurately identify histopathological variations?

Well-trained expertise in any histopathologic area is still the gold standard. What we’re seeing from this study is that certain types of histology are more easily learned by the computer than others, and that there’s room for improvement in some. However, for a first [attempt, these data are] exciting.

What are the downstream implications of this research? Who will benefit most from this technology?

The implication here is that this technology will probably be more available to more remote institutions globally that [don’t have access to quick] turnaround by an expert in a given type of rare tumor. [Instead of] sending the specimens to [an expert’s] office and having them review them, [these institutions will] likely be able to go online and [verify these specimens in a relatively short amount of time].

Although the expertise of histopathologists will never be replaced, this technology will broaden availability for more remote institutions to receive timely turnaround on data which can then be verified by an expert at a later date. [Additionally, in institutions that do have expert resources available], this technology will probably work as an internal control. However, the gold standard will always be the expert histopathologist.

How do you ultimately see this technology evolution affecting the dynamic a pathologist has with a medical oncologist?

Diagnostic radiology [is another area in which] increasingly more machine learning [is being used for] pattern recognition and imaging. This is a way to free up the intellectual bandwidth of histopathologists, radiologists, and other diagnostic specialists to be able to do more analytical work beyond just nailing the diagnosis.

[With this technology, they can] potentially look at some of the deep sequencing pathways and treatment implications as they [define tumor types]. They can also see which pathways are potentially affected in [specific cancers]. As some of the more superficial histological categorization can be done by a computer, the histopathologist has time to ask some more interesting intellectual questions about targeted therapy.

What main message would you like to leave colleagues with regarding the benefits of machine learning in histopathology?

This research is an example of why we need to carefully embrace technology and not be threatened by it, so we can enable better health care, more accessible health care, and more freedom for providers to do more intellectually rigorous endeavors, rather than some of the more rote aspects of what we do as clinicians. This technology should not be intimidating to us. Instead, it should be seen as a tool to make us more able to help our patients and enjoy our jobs.


  1. Frankel AO, Lathara M, Shaw CY, et al. Machine learning for rhabdomyosarcoma histopathology. Mod Pathol. Published online May 16, 2022. doi:10.1038/s41379-022-01075-x
Related Videos
R. Lor Randall, MD, FACS
Expert on urothelial carcinoma
Experts on urothelial carcinoma
R. Lor Randall, MD, FACS
Marcin Dzienis, MD
Nicholas Stollenwerk, MD, of UC Davis Comprehensive Cancer Center
Michael J. Wagner, MD, explains how immunotherapy may be a promising treatment approach for patients with angiosarcoma.
Michael J. Wagner, MD, discusses molecular pathways that are becoming relevant in angiosarcoma.
Michael J. Wagner, MD, discusses known risk factors for developing angiosarcoma.
Michael J. Wagner, MD, discusses the standard of care for patients with angiosarcoma.
Related Content