An investigation into the failure trends of an AI tool for the detection of tumour margins

Breast cancer is a global health challenge, affecting millions of women annually. In 2020, it was the most common cancer in 109 countries, including Canada. Breast-conserving surgery (BCS), or lumpectomy, is the standard of care for early-stage breast cancer. The goal of BCS is to remove malignant tissue while preserving healthy tissue but achieving tumor-free margins remains a significant challenge. Permanent histopathology, the gold standard for margin assessment, typically takes 2–5 days, leading to reoperations in over 20% of cases due to positive margins.

Perimeter Medical Imaging AI, a Toronto-based medical device company founded in 2013, has developed an OCT device with an embedded AI decision support module. This tool provides real-time feedback on tumor margins during surgery, potentially reducing reoperation rates. However, the AI module’s performance is not uniform across all patient subgroups and device conditions. A retrospective efficacy study of the AI tool has been published in a peer-reviewed journal, but further investigation is needed to understand its limitations and improve its reliability.

This project will focus on identifying trends in the AI tool’s failure modes, including false positives and false negatives, across different demographics, disease types, and device variability. By addressing these limitations with changes to the annotation or training data composition, we aim to enhance the tool’s accuracy and usability, ultimately improving patient outcomes.

Faculty Supervisor:

Isabelle Rao

Student:

Partner:

Perimeter

Discipline:

Engineering

Sector:

Biotechnology; Artificial Intelligence

University:

University of Toronto

Program: