Researchers have developed an artificial intelligence model that can predict in real time whether surgeons removed all cancerous tissue during breast cancer surgery by examining mammograms of the removed tissue. The model performed as well as or better than human doctors.
The preferred treatment for early-stage breast cancer is breast-conserving surgery, or partial mastectomy, combined with radiation therapy. All cancerous breast tissue must be removed during surgery to prevent the cancer from coming back. The method of inspection is to check the outer edge of the resected tissue to ensure that it does not contain cancer cells, which is the "negative margin".
Mammography of tissue (specimen mammography) is a widespread means of ensuring negative margins because it can be performed in the operating room and provides immediate feedback. However, imaging of the breast specimen may not be accurate, and if cancer cells are later found, further surgery may be needed to remove more tissue.
Researchers at the University of North Carolina (UNC) School of Medicine have developed an artificial intelligence model that can predict in real time whether cancerous tissue has been completely removed during breast cancer surgery.
Kristalyn Gallagher, one of the corresponding authors of the study, said: "Some cancers you can feel and you can see, but we can't see the tiny cancer cells that may be present at the edges of the removed tissue. Other cancers are completely microscopic. This artificial intelligence tool will allow us to more accurately analyze surgically removed tumors in real time and increase the chance of removing all cancer cells during surgery. This will prevent patients from having to undergo a second or third surgery."
To "teach" the AI model what negative and positive margins look like, the researchers used 821 specimen mammography images taken immediately after resection and matched with pathologists' final specimen reports. More than half (53%) of the images were edge positive. They also fed the model patient demographics such as age, race, tumor type and tumor size.
They found that the artificial intelligence model had a sensitivity of 85%, a specificity of 45%, and an area under the receiver operating characteristic curve (AUROC) of 0.71. Sensitivity measures the model's ability to detect positive instances, while specificity measures the proportion of true negative instances that the model correctly identifies. AUROC measures the overall performance of the model, providing a value between 0 and 1, where 0.5 represents a random guess and 1 represents perfect performance.
Compared to the accuracy of human interpretations, AI models perform as well as or better than humans, researchers say. To put this into perspective, previous studies have found that the sensitivity of imaging breast specimens ranges from 20% to 58%, and the AUROC ranges from 0.60 to 0.73.
"It's interesting to think about how AI models can use computer vision to support decision-making by doctors and surgeons in the operating room," said Kevin Chen, first author of the study. "We found that AI models were as good as or even slightly better than humans at identifying positive edges."
The model helps identify margins in patients with denser breasts. On a mammogram, dense breast tissue and tumors both appear bright white, making it difficult to distinguish healthy tissue from cancerous tissue.
The researchers say their AI model could be used in hospitals with fewer resources, such as specialist surgeons, radiologists or pathologists, to make quick, informed decisions in the operating room.
"It's like providing an extra layer of support to hospitals that may not have existing expertise," said co-corresponding author Shawn Gomez. "Rather than making best guesses, surgeons can be supported by models trained on hundreds or thousands of images and get immediate surgical feedback to make more informed decisions."
The AI model is still in its early stages, and the researchers will continue to train it with more mammograms to improve its accuracy in identifying edges. This model requires further research and validation before being applied to clinical applications.
The study was published in the journal Annals of Surgical Oncology.