AI identifies new high-risk subtype in endometrial cancer
In a recent study published in Nature Communications, a team of researchers used artificial intelligence (AI) to classify histopathological images and differentiate between endometrial cancer subtypes. The tool identified a subtype of endometrial cancer known as NSMP or No Specific Molecular Profile, which is characterized by aggressive disease and low survival rates.
Background
Endometrial cancer can be classified into four subtypes, each with very different therapeutic implications and prognoses.
The classification of these subtypes has thus far been based on inadequate clinicopathological parameters with suboptimal reproducibility, which has had a direct impact on the management of cancer.
Inconsistent histotype and grade assignment for the tumors has resulted in inaccurate risk assessment, leading to either over-treatment or inadequate treatment leading to recurrence and even death.
The Cancer Genome Atlas project showed that exome and whole genome sequencing and microsatellite instability assays can be used to stratify endometrial cancers into four prognostic subtypes based on the predominant genetic mutations.
Additionally, the development of AI tools with deep learning models is being applied increasingly in medical fields to process large amounts of image or text data. This data is then used to identify potential biomarkers and improve pathological diagnoses of cancers.
About the study
In the present study, the researchers built an AI-based image classification tool using deep-learning features that analyzed histopathological images of hematoxylin and eosin-stained slides to distinguish between the two endometrial cancer subtypes NSMP and p53 abnormal or p53abn.
In a previous study, the researchers had developed a molecular classification system for endometrial cancer that was easily applicable in clinical situations. This system divided endometrial cancer into four subtypes.
The first was the POLE mutant subtype, in which the gene involved in deoxyribonucleic acid (DNA) proofreading and repair—DNA polymerase epsilon or POLE—contained pathogenic mutations.
The second subtype was the mismatch repair deficient subtype or MMRd, in which immunohistochemistry-based diagnostic tests revealed an absence of key proteins involved in mismatch repair.
The third subtype was also diagnosed using immunohistochemistry analyses and was characterized by abnormalities in the p53 tumor suppressor protein.
The last subtype, NSMP, was diagnosed by eliminating all diagnostic characters of the other three subtypes due to the absence of any defining features.
Here, the researchers used AI-based image classification to analyze the histopathological features and distinguish between the subtypes NSMP and p53abn.
They hypothesized that a subset of patients within the subtype NSMP have tumors that are histologically similar to the tumors seen in patients within the p53abn subtype, and the application of deep-learning models to assess the hematoxylin and eosin-stained slides would help identify this subset.
For this study, the researchers used hematoxylin and eosin-stained tissue sections from hysterectomies conducted on endometrial cancer patients with the p53abn or NSMP subtypes.
The study used a discovery cohort consisting of 368 patients, and the findings were validated using two independent cohorts of 614 and 290 patients.
The researchers also conducted shallow whole-genome sequencing of representative samples from both subtypes and p53abn-like NSMP samples from the validation cohort. This data was used for analysis of copy number profiles, and gene expression profiles.
Results
The study found that the AI-based analysis of histopathological images successfully identified a subset of patients within the NSMP subtype that showed significantly lower survival rates and had a more aggressive form of cancer.
This subset consisting of aggressive tumors made up almost 20% of the NSMP tumors, constituting 10% of all endometrial cancers.
The results suggested that clinicopathological features, immunohistochemistry tests, next-generation sequencing molecular markers, and gene expression profiles might still be unable to distinguish between p53abn subtypes and these p53abn-like NSMP cases.
The deep learning model also identified tumors having tumor protein TP53 mutations even though the immunostaining for p53 was normal, which would have otherwise been a false negative based on immunohistochemistry classification.
The AI-based tool could identify the NSMP subsets with more aggressive p53abn-like cancer even when the pathological and molecular features could not predict the inferior survival outcomes.
The shallow whole-genome sequencing analysis showed that this subset of NSMP cases showed a higher proportion of altered and unstable genomes similar to the subtype p53abn but with a lower level of instability.
The findings also provided evidence of histopathological differences in this subset despite the lack of pathological or immunohistochemical distinctions with the NSMP subtype.
Conclusions
Overall, the findings indicated that the AI-based image classifier was able to distinguish between subsets of endometrial cancer patients and detect a subset with significantly inferior survival outcomes.
The researchers believe that this AI-based tool can easily be incorporated into the clinical diagnostic process to scan histopathological images routinely.
Furthermore, with additional refinement, this AI-based tool could potentially replace the more time-consuming and expensive method of molecular marker-based diagnosis.
