CHICAGO – A group of researchers from the Instituto Tecnológico Universitario Buenos Aires, Argentina (ITBA) and the Instituto Privado de Oncología Alexander Fleming Buenos Aires, Argentina have developed an artificial intelligence-based tool for predicting microsatellite instability and DNA mismatch repair (MSI/MMR) from digital pathology slides.
The team used an existing computational tool called clustering-constrained-attention multiple instance learning (CLAM), a "deep learning method that utilizes weak supervision and attention-based learning" to determine which regions were most relevant for tumor classification in whole-slide images, according to research presented on Monday at the American Society of Clinical Oncology's annual meeting.
CLAM was designed for tumor subtyping tasks that are discernible to the human eye, but the researchers adapted it to detect a molecular phenomenon: MSI/MMR, they wrote in the abstract. Julieta Chirkes, a bioengineer at ITBA who helped develop the model, said in an interview that processing these images is "very hard," so the team had to use a technology that already exists. In using CLAM, the team fine-tuned it to fit their purpose and prioritized sensitivity over specificity.
"We always try to make sure our sensitivity was high, since a false negative would be the most dangerous thing, while a false positive, although not ideal, is not as bad," Chirkes said.
Maria Pombo, head of the immunohistochemistry laboratory at the Instituto Privado de Oncología Alexander Fleming Buenos Aires, said that molecular testing for MSI/MMR is difficult to do for cancer patients in Argentina due to cost and time constraints. However, the adoption of digital pathology can allow for the testing of many people because it is "cheap and it's fast."
To train the model and internally evaluate it, the team used images from the Cancer Genome Atlas and the Clinical Proteomic Tumor Analysis Consortium for endometrial cancer, as well as images from TCGA for colorectal and gastric cancers. For external validation, the researchers used archived histological samples with MSI/MMR status from a donor laboratory in Argentina.
When tested on an internal cohort of 58 samples for colorectal cancer and gastric cancer, the tool developed by Pombo and her colleagues had sensitivity of 76.9 percent, specificity of 78.1 percent, and an area under the receiver operating curve of 76.3 percent. For the 38 endometrial cancer samples, the sensitivity was 100 percent, while specificity was 68 percent, and the area under the receiver operating curve was 92 percent.
In the external cohort, which included 109 samples, both models demonstrated a sensitivity of up to 90 percent when recognizing positive cases. However, the specificity was 45 percent, meaning "it was not possible to accurately predict [microsatellite stable] samples in all three types of tumors," the researchers wrote.
They noted that "preanalytical changes play an important role in obtaining accurate results with scanners and digitalization."
Through using a deep learning model, pathologists can separate out patients who will likely respond to immunotherapy from those who will not without relying solely on traditional techniques like immunohistochemistry, PCR, and next-generation sequencing, Pombo said. Not everyone in Argentina has access to molecular testing options, but using digital pathology images can allow that testing to reach all patients regardless of location or resources, she added.
By giving patients access to their MSI/MMR test results, "you give the opportunity to these patients to have the diagnosis, the treatment, and the prognosis," she said.
The only requirements are H&E-stained slides, a whole-slide scanner, and the software tool developed by the team, Pombo said. While many laboratories and hospitals in Argentina don't have whole-slide scanners yet, Chirkes said they hope their research shows how valuable the information gathered from that equipment can be and encourages other hospitals and labs to adopt digital pathology tools.
Pombo added that digital pathology is "brand new" in Argentina.
In addition, the researchers developed a related platform that allows anyone in a lab to use the tool they've developed without computer or programming expertise, said Ana Gorodisch, a fellow bioengineer from ITBA who also helped develop the model. It only requires the uploading of the image and pressing a button to receive results. "The idea is to make the diagnostic available to everyone," she said.
With the platform, the user can also see a heat map that shows which tissue regions are most significant for the prediction of MSI/MMR, said Andrea Erbetti, another bioengineer from ITBA.
The researchers plan to further optimize the model by training it on more images from Argentina to boost its performance and make it more specific to the country before it can be launched clinically or commercially, but the team hopes it can be used for first-line screening before further molecular testing is conducted, Gorodisch added. The goal is to reduce the number of patients who receive unnecessary molecular testing, thereby reducing cost and time spent.
When a patient receives a positive result from the model, it "justifies asking for an additional test," Chirkes added.
The researchers are also determining the best way to fit the model into the workflow of clinic, Gorodisch said.
The researchers also hope to use other artificial intelligence techniques to make a more complex model that can provide additional data, since the existing model is relatively simple and "AI is a field that is growing each day, every second," Gorodisch said. "We need to stay on the frontier of what's happening."
"We can use the one that we have, but we want to improve the model every day in order to have" the best performance, Pombo said.