August 13 , 2024 | Researchers at The Australian National University (ANU) have developed a deep learning algorithm, known as DEPLOY, to classify central nervous system tumors into 10 major categories from histopathology images. The goal is to optimize treatment with a faster, more accessible diagnostic approach than the state-of-the-art practice of DNA methylation-based profiling, which can take weeks, according to Danh-Tai Hoang, Ph.D., research fellow at the Biological Data Science Institute of ANU.
In research that was published recently in Nature Medicine (DOI: 10.1038/s41591-024-02995-8), DEPLOY demonstrated 95% overall accuracy and 91% balanced accuracy. While the former “treats every patient equally,” explains Hoang, the latter “gives equal considerations to every class.”
The model was also able to provide a “more clinically relevant” diagnosis for 309 particularly hard-to-classify samples than what pathologists initially identified as the cause. “Some brain tumors have similar histological features to others, making them difficult to distinguish from each other,” says Hoang. “Pathologists mostly provided high-level diagnoses for the 309 samples... [while] DEPLOY was able to classify them with higher resolution” into one of the 10 known brain subtypes.
DEPLOY was trained and validated on large datasets of approximately 4,000 patients, 1,796 for training purposes and 2,156 for testing, he says. Unlike the traditional end-to-end approach that classifies tumor types from histopathology images directly, DEPLOY integrates three distinct components.
The first component classifies tumors directly from histopathology images (direct model), continues Hoang. The second initially predicts DNA methylation from histopathology images, and subsequently classifies tumors based on the predicted methylation (indirect model). The third classifies tumors directly from patient demographics (gender, age, and biopsy location).
“Interestingly and surprisingly, the performance of the indirect model exceeds that of the direct model,” he notes. “More importantly, the integrated model outperforms each of the individual models.”
The take-home message from the latest study here is twofold, says Hoang. “First, in terms of biology, pathology images contain rich information.” In the paper, DEPLOY was able to accurately predict DNA methylation from slide images.
“Second, in terms of machine learning, the indirect model that initially predicts DNA methylation and subsequently classifies brain tumors from the predicted DNA methylation outperformed the direct model that classifies brain tumors directly from images without the intermediate step of DNA methylation prediction,” he says.
Despite its encouraging performance, DEPLOY won’t be replacing pathologists, “at least at the moment,” says Hoang. The hope is that DEPLOY could serve as a complementary tool, providing confidence in a pathologist’s diagnosis if it is consistent with the one provided by DEPLOY, or to suggest review and possibly further tests if they differ.
The same model architecture could be used for other cancer types, he adds. “The most important requirement is a dataset that contains sufficient samples for model training.”
Hoang and his team are now developing some new model architectures in addition to collecting more data samples. “In the future, subject to careful regulatory approvals, we would like to develop a web server and a mobile app,” he says. “This will enable pathologists worldwide to upload tumor slide images to our model platform, obtaining an AI-based ‘second opinion’ diagnosis for their possible further consideration.”