AI model simultaneously detects multiple genetic colorectal cancer markers in tissue samples

A multicenter study has analyzed nearly 2,000 digitized tissue slides from colon cancer patients across seven independent cohorts in Europe and the US. The samples included both whole-slide images of tissue samples and clinical, demographic, and lifestyle data.

The researchers have developed a novel “multi-target transformer model” to predict a wide range of genetic alterations directly from routinely stained histological colon cancer tissue sections. Previous studies were typically limited to predicting single genetic alterations and did not account for co-occurring mutations or shared morphological patterns.

The model detects genetic alterations and resulting tissue changes in colorectal cancer directly from tissue section images. This could enable faster and more cost-effective diagnostics in the future. For the development, validation, and data analysis of the model, experts in data and computer science, epidemiology, pathology, and oncology worked closely together.

The study has been published in the journal The Lancet Digital Health.

“Earlier deep learning models and analyses of the underlying tissue alterations have generally focused on only a single mutation at a time. Our new model, however, can identify many biomarkers simultaneously, including some not yet considered clinically relevant. We were able to demonstrate this in several independent cohorts. We also observed that many mutations occur more frequently in microsatellite-instable tumors (MSI),” explains Marco Gustav, M.Sc., first author of the study and researcher at EKFZ for Digital Health at TU Dresden.

Certain types of colorectal cancer can be classified based on microsatellite instability (MSI). Microsatellites are short, repetitive DNA sequences spread throughout the genome. In cancer, MSI can occur when these sequences become unstable due to defects in the DNA repair system. MSI is an important biomarker for identifying patients who may benefit from immunotherapy.

“This suggests that different mutations collectively contribute to changes in tissue morphology. The model recognizes shared visual patterns, rather than independently identifying individual genetic alterations,” he adds.

The researchers demonstrated that their model matched and partly exceeded established single-target models in predicting numerous biomarkers, such as BRAF or RNF43 mutations, and microsatellite instability (MSI) directly from pathology slides.

The pathological expertise required to assess tissue changes from histological slides was provided by experienced medical specialists. Dr. Nic Reitsam from the University Hospital Augsburg played a key role in the study.

Highlighting the study’s significance, Jakob N. Kather, Professor of Clinical Artificial Intelligence at the EKFZ for Digital Health at TU Dresden and senior oncologist at the NCT/UCC of the University Hospital Carl Gustav Carus Dresden, says, “Our research shows that AI models can significantly accelerate diagnostic workflows.

“At the same time, these methods provide new insights into the relationship between molecular and morphological changes in colorectal cancer. In the future, this technology could be used as an effective pre-screening tool to help clinicians select patients for further molecular testing and guide personalized treatment decisions.”

The research team now plans to extend this approach to other types of cancer.

A multicenter study has analyzed nearly 2,000 digitized tissue slides from colon cancer patients across seven independent cohorts in Europe and the US. The samples included both whole-slide images of tissue samples and clinical, demographic, and lifestyle data.

The researchers have developed a novel “multi-target transformer model” to predict a wide range of genetic alterations directly from routinely stained histological colon cancer tissue sections. Previous studies were typically limited to predicting single genetic alterations and did not account for co-occurring mutations or shared morphological patterns.

The model detects genetic alterations and resulting tissue changes in colorectal cancer directly from tissue section images. This could enable faster and more cost-effective diagnostics in the future. For the development, validation, and data analysis of the model, experts in data and computer science, epidemiology, pathology, and oncology worked closely together.

The study has been published in the journal The Lancet Digital Health.

“Earlier deep learning models and analyses of the underlying tissue alterations have generally focused on only a single mutation at a time. Our new model, however, can identify many biomarkers simultaneously, including some not yet considered clinically relevant. We were able to demonstrate this in several independent cohorts. We also observed that many mutations occur more frequently in microsatellite-instable tumors (MSI),” explains Marco Gustav, M.Sc., first author of the study and researcher at EKFZ for Digital Health at TU Dresden.

Certain types of colorectal cancer can be classified based on microsatellite instability (MSI). Microsatellites are short, repetitive DNA sequences spread throughout the genome. In cancer, MSI can occur when these sequences become unstable due to defects in the DNA repair system. MSI is an important biomarker for identifying patients who may benefit from immunotherapy.

“This suggests that different mutations collectively contribute to changes in tissue morphology. The model recognizes shared visual patterns, rather than independently identifying individual genetic alterations,” he adds.

The researchers demonstrated that their model matched and partly exceeded established single-target models in predicting numerous biomarkers, such as BRAF or RNF43 mutations, and microsatellite instability (MSI) directly from pathology slides.

The pathological expertise required to assess tissue changes from histological slides was provided by experienced medical specialists. Dr. Nic Reitsam from the University Hospital Augsburg played a key role in the study.

Highlighting the study’s significance, Jakob N. Kather, Professor of Clinical Artificial Intelligence at the EKFZ for Digital Health at TU Dresden and senior oncologist at the NCT/UCC of the University Hospital Carl Gustav Carus Dresden, says, “Our research shows that AI models can significantly accelerate diagnostic workflows.

“At the same time, these methods provide new insights into the relationship between molecular and morphological changes in colorectal cancer. In the future, this technology could be used as an effective pre-screening tool to help clinicians select patients for further molecular testing and guide personalized treatment decisions.”

The research team now plans to extend this approach to other types of cancer.