Anomaly detection using intraoperative iKnife data: a comparative analysis in breast cancer surgery

Area of Science:

• Oncology
• Surgical Technology
• Artificial Intelligence

Background:

• Intraoperative margin assessment is vital for successful breast-conserving surgery.
• The Intelligent Knife (iKnife) analyzes surgical smoke for real-time margin evaluation.
• Current AI models require costly, time-consuming labeled ex-vivo datasets.

Purpose of the Study:

• To explore anomaly detection models for margin assessment using unlabeled intraoperative spectra.
• To reduce reliance on labeled ex-vivo datasets for AI model training.
• To assess the feasibility of anomaly detection for intraoperative margin identification.

Main Methods:

• Collected iKnife spectra intraoperatively from 15 breast cancer surgeries.
• Trained four anomaly detection models (iForest, OCPCA, GODS, KGODS) using intraoperative data, with and without healthy ex-vivo data.
• Evaluated performance via cross-validation on labeled ex-vivo samples and intraoperative feasibility testing.

Main Results:

• Anomaly detection models achieved balanced accuracies between 70-81% using intraoperative data alone.
• OCPCA trained solely on intraoperative data yielded 81% balanced accuracy, 90% sensitivity, and 72% specificity.
• OCPCA demonstrated intraoperative feasibility, correctly identifying tumor breach with one false positive.

Conclusions:

• Anomaly detection models, especially OCPCA, can identify positive breast cancer margins without labeled ex-vivo data.
• These models present a viable low-resource alternative to supervised classifiers for intraoperative margin assessment.
• The approach supports intraoperative label generation and semi-supervised training, enhancing clinical deployment.