?Background: Advanced stage breast cancer rates are considerably higher in Hawaii (15%) and the USAPI (>50%) than in the continental US (9%). This same trend is observed in many remote, rural, and low- and middle-income countries (LMICs) worldwide. Histopathology is the current standard of care for breast cancer. However, extensive infrastructure is needed for histology. The World Health Organization’s Global Breast Cancer Initiative (GBCI) states completion of diagnosis, from first symptoms to pathological confirmation, be within 60 days or mortality will increase. Fine needle aspiration biopsy (FNAB) is less invasive than core biopsies and can be evaluated on site, but requires expert pathological assessment. In this work, we ask if artificial intelligence (AI) combined with whole-slide cytology images from breast FNAB can accurately diagnose breast cancer. Using AI reduces the training needed to provide cytology services locally by offloading cellular adequacy and preliminary diagnosis to the AI model, reducing telepathologist workload.

Objectives: Develop and validate AI solutions for cytopathology that: (1) assess specimen cellular adequacy at the point of care, (2) support onsite preliminary interpretation using rapid staining, and (3) enable definitive clinical diagnosis using standard-of-care staining and ancillary testing workflows. Our approach is to determine if histopathology foundation models can be adapted for cytopathology cell smears.

Methods: The CytoBreastDx model was developed from 482 digitized images of breast FNAB smears, collected from the Wisconsin Diagnostic Breast Cancer Dataset (Street, Wolberg et al. 1993). Because of the low volume of training data, we use Virchow2 (Zimmermann, Vorontsov et al. 2024) as a feature extractor. Virchow2 is a pathology foundation model. We develop and test logistic regression, XGBoost, and decision trees, with and without dimensionality reduction techniques.  Patients were split into a development set (70%), and a completely held-out test (30%). Test set AUROC values were used to evaluate the ability to classify cytology images.

Results: Virchow2 produced 2,560 features from its penultimate layer. The most successful model was logistic regression with L1 regularization which achieved an AUROC of 0.98 on the held-out test set. Logistic regression after principal components analysis, XGBoost, and the decision tree achieved AUROCs of 0.97, 0.97, and 0.78, respectively.

Conclusions/Implications: Accurate malignancy classifications are possible using a model trained with a low volume of cytology images. A pathology foundation model trained exclusively on histology slide images was able to extract features meaningful for cytology slides with no modality-specific finetuning. The results show that offloading of cytopathological interpretation of cytology smears to an AI model to increase access to rapid diagnosis after FNAB in remote and rural LMICs may be possible.

References: Street, W. N., W. H. Wolberg and O. L. Mangasarian (1993). Nuclear feature extraction for breast tumor diagnosis. Biomedical image processing and biomedical visualization, SPIE.
Zimmermann, E., E. Vorontsov, J. Viret, A. Casson, M. Zelechowski, G. Shaikovski, N. Tenenholtz, J. Hall, D. Klimstra and R. Yousfi (2024). "Virchow2: Scaling self-supervised mixed magnification models in pathology." arXiv preprint arXiv:2408.00738.