Abstract 171P
Background
Fine-needle aspiration cytology (FNAC) is a minimally invasive, very rapid, cost-effective, safe, and reliable modality for diagnosing breast lesions. The radiology-cytological diagnosis represents a highly accurate tool in the assessment of breast masses in structures where core-needle biopsy is not available. Our study introduces an innovative AI-based diagnostic tool to differentiate malignant and benign lesions from FNAC whole slide images (WSI).
Methods
We collected 284 Diff-Quik-stained conventional cytology slides, corresponding to 119 benign and 165 malignant (carcinomas) consecutive BI-RADS 3-5 breast masses, diagnosed in Gustave Roussy’s one-stop breast clinic over a period of 7 months. For all these cases, either a histopathological verification or an 18-month ultrasound follow-up evaluation was available. The WSI were scanned at 40X magnification, and partitioned into non-overlapping patches of 256x256 px. We applied an attention-based multiple instance learning model from representations extracted from each patch to predict malignancy. This attention mechanism provides valuable insights by identifying which patches were the most contributive to the prediction, providing some level of interpretability.
Results
Using a 10-fold cross-validation stratified by patient (80/10/10% train/val/test split) of the 284 samples, our model achieved remarkable testing performance: average Area Under the Curve (AUC) of 98.43 (95% CI: 97.53–99.32), sensitivity of 95.15 (95% CI: 91.72–98.57), specificity of 89.92 (95% CI: 79.88–99.97), Positive Predicted Value (PPV) of 93.74 (95% CI: 87.87–99.61), and Negative Predicted Value (NPV) of 93.70 (95% CI: 89.49–97.90). Among benign samples, error rates varied by lesion type, with mesenchymal and cystic lesions demonstrating lower error rates than solid and inflammatory lesions (Wilcoxon U-test, p=0.038).
Conclusions
Utilizing FNAC, a readily available technique, our AI tool effectively diagnosed breast cancer from WSI. In particular, the high sensitivity and PPV of our model suggest a potential application in clinical practice in regions with limited healthcare facilities, offering new avenues for low-cost and fast access to breast cancer diagnosis.
Legal entity responsible for the study
The authors.
Funding
National PRecISion Medicine Cancer Center (PRISM).
Disclosure
S. Delaloge: Financial Interests, Institutional, Advisory Board: Novartis, Sanofi, Gilead; Financial Interests, Institutional, Invited Speaker: Pfizer, Roche Genentech, BMS, Sanofi; Financial Interests, Institutional, Advisory Board, ad board: Besins Healthcare; Financial Interests, Institutional, Invited Speaker, ESMO symposium: Gilead; Financial Interests, Institutional, Advisory Board, scientific board: Elsan; Non-Financial Interests, Member of Board of Directors, Société Française de Sénologie et Pathologie Mammaire: SFSPM; Non-Financial Interests, Principal Investigator, H2020 funding: European Commission. F. André: Financial Interests, Personal, Advisory Board: Lilly FRANCE; Financial Interests, Institutional, Advisory Board: AstraZeneca, Daiichi Sankyo, Roche, Lilly, Pfizer, OWKIN, Novartis, Guardant Health, N-Power Medicine, Servier, Gilead, Boston Pharmaceutics; Financial Interests, Institutional, Research Grant: AstraZeneca, Lilly, Novartis, Pfizer, Roche, Daiichi, Guardant Health, OWKIN. All other authors have declared no conflicts of interest.