Skip to content
Prof. Dr. Mustafa Özdoğan · Medical Oncologist

Much More Accurate Cancer Classification with Machine Learning

Much More Accurate Cancer Classification with Machine Learning

Many types of cancer are commonly classified according to how their cells look under a microscope. This method, called pathological examination, has been performed by pathologists for nearly 200 years. Although pathologists are highly specialized and base their assessments on numerous criteria and algorithms, diagnoses have subjective features: they may vary somewhat between observers. Using machine learning , one of today’s artificial intelligence technologies, a brain tumor study examined molecular patterns and achieved more objective and precise cancer diagnoses.

The most fundamental aspect of treating a disease is identifying it correctly. Current developments in artificial intelligence-based image processing allow precise diagnostic studies independent of subjective results. Capper and colleagues are working on these machine learning-based technologies and have presented a new classification of brain tumors based on molecular patterns.

Today, many tumors are analyzed and defined by molecular features as well as histological characteristics. In 2016, the World Health Organization (WHO) recommended combining histology and molecular information in integrated diagnostic approaches. Nevertheless, diagnoses based predominantly on histology remain inadequate for many rare tumor types because of a lack of molecular identifiers. Histological diagnoses also face numerous challenges, including possible cellular variations within tumors that form mosaics of cells carrying different genetic changes, and the fact that many different brain tumors can share similar histological features. Questions remain about how well histological similarity reflects tumor similarity, considering that tumors with similar histology may progress differently and tumors with different histology may progress similarly.

An important advance in histological analysis is the expansion of computational tools enabling machine learning processes to analyze histological data. In this approach, using a dataset of tumor sample images classified by a physician, a computer is trained. The computer uses classification information to develop its own pattern-recognition criteria for identifying tumor types. However, difficulties arise if clearly defined diagnostic criteria for certain tumors are lacking or different tumor types are histologically indistinguishable.

Capper and colleagues decided to focus on molecular information whose classification does not require complex visual assessment, and on epigenetic changes. They applied a machine learning approach to tumor classification based on changes in DNA methylation (the addition of methyl groups to DNA) and compared these diagnoses with those made by pathologists using histological analyses.

DNA methylation is a type of modification known as an epigenetic change. This category of changes does not alter the DNA sequence but can affect gene expression or a cell’s fate. The role of DNA methylation and other epigenetic changes in cancer is becoming increasingly apparent. In many cancers, the genome-wide pattern of epigenetic changes, known as the epigenome, can be substantially altered. For example, mutations in the IDH1 or IDH2 genes in glioma brain tumors cause genome-wide disruption of DNA methylation patterns that can be associated with specific clinical outcomes.

Previous studies have highlighted the diagnostic advantages of profiling DNA methylation for specific brain tumors, because epigenome-wide analysis of DNA methylation offers an unbiased diagnostic approach. Nevertheless, routine epigenome-wide methylation profiling remains relatively uncommon in clinical diagnosis for several reasons: cost; sample requirements; lack of personnel with the required data analysis expertise; and the question of whether the findings affect the clinical treatments used. However, progress is being made. For example, techniques are available for using DNA extracted from the most commonly chemically preserved tumor tissue on glass slides, known as formalin-fixed, paraffin-embedded (FFPE) samples.

The authors supplied computers with genome-wide methylation data for samples representing almost every WHO-classified brain tumor type. The computers used supervised machine learning on samples classified by pathologists, as well as information from unsupervised machine learning, to recognize methylation patterns.

After training, the computer could classify tumors into 82 distinct classes based on specific methylation profiles. Only 29 of these corresponded to a specific tumor type defined within the 29 represented subcategories of WHO-defined tumor types.

Nevertheless, the most interesting discoveries by Capper and colleagues involved tumor classifications that grouped histologically similar tumors, such as classifications spanning multiple WHO tumor types or tumor types that did not match WHO groupings. Such discoveries may provide information about tumor similarities independent of histology and help develop treatment options or diagnostic tools.

The authors used the computer to classify 1.104 test cases of tumors diagnosed by pathologists using standard histological or molecular techniques. In %60.4 of these cases, computer-based classification was identical to the pathologist’s classification. In %15.5, the computer and pathologist assigned the same tumor type, but the computer could additionally assign a subclass. In %12.6 of test cases, the computer diagnosis disagreed with the pathologist’s diagnosis.

Remarkably, more detailed analysis of these cases—for example, gene sequencing —resulted in %92.8 of these unmatched tumors being reassigned from the original clinical diagnosis to the computer-based classification. Moreover, %71 of the reclassified tumors were assigned a different tumor grade, a reclassification potentially affecting prognosis (disease course) or treatment. The remaining test cases (% 11.5) could not be classified by the computer. Additional computational analysis indicated that one-third of tumors in this group might represent rare tumors for which the computer had encountered enough samples to establish a classification grouping.

Is Capper and colleagues’ approach comparable to standard cancer diagnosis? Given advantages such as low cost per sample, does it represent a possible future standard for tumor diagnosis? If so, will this mean the end of histological analyses?

Obtaining a comprehensive molecular profile of a tumor sample is certainly useful, particularly when combined with microscopic examination, and may be the way forward as medical treatments become more personalized to the characteristics of an individual’s tumor. For now, however, histology remains indispensable for disease classification because standard approaches to sample preservation and microscopic examination offer the most accessible and universal entry point into routine diagnostic workflows in clinical laboratories worldwide. A disease can manifest through both molecular and cellular changes; therefore, an approach combining molecular analysis and visual examination may strengthen diagnostic capabilities.

Routine, widespread use of the platform developed by Capper and colleagues may currently be impractical for many laboratories, so its most immediate application would be evaluating cases with ambiguous histological features. However, if this machine learning approach is refined and expanded, it may replace diagnostic microscopic examination in some cases.

Sources

Derek Wong, Stephen Yip.
Machine learning classifies cancer.
Nature, News and Views, 14 March 2018

Appointment request

DROZDOGAN AI

Explore this article together

Ask about the article and read answers alongside published sources from our site.

DROZDOGAN AI

Drozdogan Assistant

Much More Accurate Cancer Classification with Machine Learning

Ask about the article and read answers alongside published sources from our site.

Conversation and privacy information

Drozdogan Assistant answers this conversation. Authorized site administrators may review it for quality. Do not share personal information.

Replies use OpenAI infrastructure. Questions and replies are processed by that service; conversation records are stored on our site. Do not share identifying, contact or personal health information.

0 / 1000

For information only; no personal diagnosis or treatment advice.

Related articles

Explore other articles on this topic.

Stay informed

Receive new articles and scientific updates by email.

Google Analytics runs on public pages independently of this window and may use analytics cookies. Your choice here controls external services only.

Cookie policy