Artificial intelligence as a negative predictive tool for breast cancer postoperative recurrence

Abstract

Background

Mammography alone is an ineffective method for breast cancer surveillance and diagnosing cancer recurrence. The aim was to evaluate the ability of artificial intelligence (AI) to read digital mammograms as an additive tool to exclude recurrence in the operative bed of known breast cancer patients following the different surgical procedures.

Methods

We used a retrospective cohort study of post-surgery mammograms (n = 577). Imaging was performed within 6 months after the surgery or more. The AI solution used to read mammograms (AI-MMG) provided a targeted heat map of the operative bed, which was supported by a decision likelihood score percentage of cancer recurrence. The reference for suspicious or malignant-looking abnormalities (n = 62, 12.3%) was diagnosed by biopsy. A clear operative bed and benign-looking changes (n = 442) were confirmed by ultrasound characterization patterns and one year of intermittent follow-up.

Results

The AI scoring percentage for a clear operative bed ranged between 0 and 26%, with a mean of 15% ± 5.4%. Operative bed benign changes ranged from 10 to 88%, with a mean of 48.2% ± 21.2%, while malignancy recurrence ranged from 65 to 99%, with an average of 87.7% ± 10.5%. The “ROC: Receiver Operating Characteristic” curve for AI to predict cancer in the surgical bed on mammograms was 0.906. The optimum cutoff value to distinguish between benign postoperative alterations and malignancy recurrence was 56.5% (95%, CI 0.824–1.060, p value < 0.001).

Excellent agreement between AI-MMG and pathology or ultrasound results was observed, and Kappa was 0.894, p value < 0.001.

Conclusions

The use of artificial intelligence has enhanced the diagnostic performance of the postoperative mammograms to rule out recurrent malignancies in breast cancer surveillance.