AI tool helps diagnose an often-misclassified group of rare thymic tumors

A team led by researchers at the University of Chicago developed an artificial intelligence (AI) tool that could help pathologists more accurately diagnose a rare group of tumors known as thymic epithelial tumors (TETs). The work is described in a recent study published in the Annals of Oncology.

TETs are cancers arising from the thymus — a small gland situated in the upper chest that plays an important role in the immune system. In the United States, these cancers affect just 2-3 people per million each year.

“This is a very rare type of cancer, so very few people in the world are trained to diagnose and treat it,” said senior author Marina Garassino, MD, Professor of Medicine at UChicago Medicine.

Adding to that challenge is the fact that these tumors can look and behave very differently from one patient to another. Together, visual and clinical features help define five main TET subtypes. Given the aggressive nature of these cancers, accurate diagnosis is crucial for guiding treatment decisions, as misclassification may lead to patients not receiving the most appropriate care.

“This study started from my research in Italy, where I published that in non-academic centers with non-expert pathologists, there is a large discrepancy in the diagnosis of TETs of about 40%,” Garassino said.

To help close this gap, the team turned to AI and digital pathology. They trained a computational model to recognize patterns in microscope images of tumors from 119 TET patients, essentially teaching the model what different TET subtypes “look like.” The images came from The Cancer Genome Atlas Program (TCGA), a large, publicly available cancer dataset in which subtype classifications have been confirmed by expert pathologists.

The researchers then tested their model on 112 cases from the University of Chicago, with diagnoses confirmed by an expert pathologist. The tool classified TET subtypes with high accuracy overall, and was especially effective at identifying thymic carcinomas, the most aggressive subtype.

“Basically, we created a tool that — in the hands of a non-expert pathologist — is able to properly diagnose 100% of thymic carcinomas and outperform non-expert diagnoses,” Garassino explained.

The tool is freely available and is not intended to replace pathologists. Instead, its goal is to support clinicians who do not specialize in rare thymic tumors in diagnosing and treating the cancer.

“One of the biggest challenges and also the beauty of this work was bringing data scientists, pathologists, and oncologists together at the same table,” Garassino said. “It was a truly multidisciplinary effort, and we learned from each other about what we could and couldn’t do.”

The team is now working to validate their tool at a much larger scale, including data from additional cancer centers in the United States and Europe.

One remaining challenge is that their current model was trained on data that used similar laboratory and imaging procedures. Differences in how microscope slides are prepared and scanned across hospitals can affect how tumors appear in images.

“In a larger population, harmonizing these steps is the biggest challenge,” Garassino said. “So, in the future, we plan to expand the algorithm so that it can correct for such differences, which will make the tool even more widely usable.”

The study “Deep learning discriminates thymic epithelial tumors’ histological subtypes using digital pathology” was supported by grants from the National Institutes of Health, and a scholarship “Pierluigi Galli and Eurovetro Recycling SRL” from Associazione TUTOR. The study was also made possible with the support of the Department of Medicine, Section of Hematology/Oncology and Department of Pathology at The University of Chicago and the TCGA Research Network.

Additional authors include Matteo Sacco, Anna Di Lello, Alexander McGeough, Alessandra Esposito, Rishi Sharma, Aliya Husain, Qudsia Arif, Maha Elsebaie, Alexander Pearson from UChicago; James Dolezal from Geisinger Cancer Institute; Erica Pietroluongo from University of Naples Federico II; Mirella Marino from IRCCS Regina Elena National Cancer Institute.