178P Deep learning reveals spatial disorganisation of histological features in the normal breast of breast cancer patients
S. Chen, Mario Parreno-Centeno, G. Verghese, G. Booker, I. Wall, F. Mohamed, S. Arsian, Pandu Raharja-Liu, Aasiyah Oozeer, Maria Grazia D’Angelo, Rachel Barrow, Rachel L. Nelan, Marcelo Sobral‐Leite, Esther H. Lips, F. de Martino, Cathrin Brisken, Cheryl Gillett, J. Louise Jones, Sarah E. Pinder, Anita Grigoriadis · ESMO Open · 2023
Germline BRCA1/2 mutation carriers have an increased risk of developing breast cancer, however the characterisation of histological features of the normal breast for these high-risk patients has not so far been performed on large-scale image collections. Hence, deep learning-based framework is proposed to identify subvisual morphometric phenotypes in normal breast tissue of women with different risk of developing cancer. Digitised 1190 whole slide images of H&E-stained breast normal tissue images from women across different age groups were collected from 5 different biobanks, i.e., healthy donors, healthy patients derived from reduction mammoplasty, healthy patients with known germline BRCA1/2 alterations derived from risk reduction mastectomy, and the contralateral/ipsilateral tissue from breast cancer patients with known germline pathogenic BRCA1/2 alterations, totalling 282 patients. To characterise the epithelial patterns, 70 WSI were manually annotated. A convolution deep learning (DL)-network was implemented to predict the chronological age group (50y) of tiles of epithelium. K-means clustering identified specific histological age-associated features. Class activation map and cellular segmentation were used for interpretation. Our DL-based framework revealed inter and intra-variability of epithelial histological patterns in the normal breast of healthy women, underlying the physiological ageing process. The proportion of lobule types differed in normal breast tissue of 50y healthy women. In normal breast of younger women, densely packed acinar structures with a higher glandular-to-stromal ratio are predominant and display higher neighbourhood enrichment. Normal breast tissue of young breast cancer patients revealed histological tissue-ageing patterns deviating from their chronological age. Our DL-framework exposed histological patterns associated with the variation of the underlying ageing process of normal breast tissue, which may indicate early signs of cancerous in high-risk germline BRCA1/2 carriers.