AI detects breast cancer in images produced through mass-spectrometry

The main advantage of a lumpectomy for patients with early-stage breast cancer is that it conserves the maximum amount of healthy breast tissue possible. The surgery involves removing only the tumor and a margin, a thin band of healthy tissue that surrounds the cancer. When successfully performed and followed by adjunctive therapy such as radiation treatment or chemotherapy, the treatment can be just as effective as a mastectomy.

However, success hinges on what’s in that margin. If a pathologist’s analysis reveals lingering cancer cells near the margin’s outer edges, the patient will likely require a second surgery. This isn’t the odd case: an estimated 20%-25% of lumpectomies require re-excision to clear the margins. Unfortunately, because a pathologist’s analysis can take weeks to perform, any remaining cancer cells have the opportunity to grow in the interim, potentially requiring the removal of even more tissue in the second operation.

New machine learning research shows promise for speeding up this analysis and, as a result, potentially reducing the amount of tissue requiring removal in re-excisions. The research, led by Queen’s University master’s student Rachel Theriault in collaboration with Dr. Randy Ellis from Queen’s School of Computing, demonstrates how a machine learning algorithm typically used for facial recognition can be repurposed to detect cancer cells in breast tissue samples.

The algorithm analyzes metabolic patterns in DESI-MS scans – images produced through mass-spectrometry. Theriault’s research determined that the algorithm can successfully distinguish tissue from an image’s background and can detect enough metabolic difference between tissue types to separate and classify malignant, benign, and artifact tissues such as cauterized tissue and fibrous bands. When the researchers compared the algorithm’s tissue classifications to those made by pathologists on the same samples, they matched.

These results mark important progress for computational analysis of cancer tissue. Previous efforts to apply machine learning to classify breast tissues via DESI-MS scans have fallen short due to the sheer volume of metabolite-related data that needs to be accounted for. Theriault explains, “There are lots of differences between the tissues that are very hard to capture, and the cancer itself is heterogeneous, so it’s hard to determine what is cancer and what is benign. We needed a complex algorithm to solve this complex problem.”

That advanced algorithm is sparse subspace clustering, which clusters in high dimensional data and is typically used for facial recognition and video processing.

Theriault says, “We thought, if the algorithm can capture the complex patterns in video processing, it can probably capture the tissue’s heterogeneity.”

She and Dr. Ellis hypothesized that sparse subspace clustering should be able to cluster pixels based on the similarities of mass spectra (representing various ions) to identify tissue types in the same way that it clusters features like edges, shapes, or colourings in the pixels of an image to detect objects or to identify a specific face.

The confirmation that sparse subspace clustering can treat ions as features to distinguish and classify tissue types opens the door to using machine learning to expedite the analysis of tissue samples from lumpectomies. However, Theriault stresses that this remains fundamental research, and still requires significant study and development before it is ready for application in a health care setting.

“These results aren’t the end. They’re just the beginning,” Theriault says.

The next step is to determine whether sparse subspace clustering can accurately classify tissue types across samples. Ambient air pressure affects DESI-MS, and because air pressure is not a constant, scans of a sample taken at different times can produce different image results.

Research must determine whether sparse subspace clustering reliably detects metabolic patterns across such differing results.

Theriault continues this work in a master’s program at Queen’s University, where in addition to this project, she is also exploring the application of sparse subspace clustering to the analysis of skin, liver, and prostate cancers.

To support her work, Theriault was awarded a Vector Scholarship in AI by the Vector Institute, a Toronto-based institute advancing AI development and adoption across Canada, with a strategic pillar dedicated to health-related research and application.

Ultimately, Theriault’s aspiration is that her work with sparse subspace clustering can lead to real impact in a clinical setting. She says, “When we have computational strategies that can do intraoperative analysis or help guide a pathologist to do faster analysis by flagging certain samples, we hope that we can decrease the need for second surgeries.”