Abstract P5-01-01: Multiscale Deep Learning framework to capture systemic immune features in lymph nodes predictive of triple negative breast cancer outcome

Abstract

Abstract Systemic immune responses in lymph nodes (LN) convey significant prognostic value for breast cancer patients, which can inform disease progression and optimal treatment management. However, have, so far, not been assessed in large patient cohorts. We have previously shown that morphological alterations in axillary LNs, namely the formation of germinal centres (GCs) in cancer-free LNs, add prognostic value to tumour infiltrating lymphocytes (TILs) in triple-negative breast cancer patients (TNBC) for the development of distant metastasis. Extending manual assessment of LNs beyond the detection of cancer requires the integration of robust deep learning pipelines into the digital pathology workflow. Here, we propose a supervised multiscale deep learning framework named smuLymphNet to capture and quantify GCs and sinuses within LNs from digitised Haematoxylin and Eosin-stained (H&E) whole slide images (WSIs) and show good concordance compared with an inter-pathologist Dice coefficient of manual annotations from four pathologists. The smuLymphNet framework consists of (i) a detection algorithm to determine the boundaries of each LN section on the WSI, using an Otsu-based thresholding method and contouring algorithm; (ii) a supervised multiscale deep learning module for the segmentation of GCs and sinuses; and (iii) quantification of the number, size, and shape of the predicted features. We applied smuLymphNet to a total of 1,800 H&E-stained WSI of >4,000 cancer-free and involved LNs from a retrospectively collected breast cancer cohort collected at Guy’s Hospital (London, UK) from 177 patients (122 N+) enriched for the triple-negative phenotype. A subset of 114 WSI and five breast cancer LN WSIs from each Barts Hospital (London, UK) and Tianjin University Hospital (Tianjin, China) were used to train and evaluate the supervised deep learning module. For training Fully Convolutional Networks (FCNs), WSIs manually annotated for both GCs and sinuses formed a ground-truth set and three FCNs were implemented: (i) a standard U-Net architecture; (ii) a U-Net model with an attention gate mechanism; and (iii) a multiscale-U-Net network (MS U-Net) that encodes, in parallel, a feature representation of the image at multiple resolutions. The MS U-Net achieved the best performance with an average dice score of 0.86 for GCs and 0.74 for sinuses. In comparison, the average dice score amongst four pathologists assessing 24 LN WSI for GCs and sinuses was 0.67 and 0.61, respectively, demonstrating the robustness of the smuLymphNet framework. To establish associations between morphometric immune features and patients’ outcomes, we assessed smuLymphNet captured GCs and sinuses from 686 WSIs from 96 TNBC patients with extensive longitudinal outcome data. We found significant morphological differences in involved and cancer-free LNs between N0 and N+ patients, with the latter displaying larger GCs with more irregular shapes, especially in their involved LNs. Moreover, in alignment with our previously published studies, our multiscale smuLymphNet framework recapitulated and extended the prognostic value of the assessment of GC formation in TNBC N0 patients. We further revealed, for the first time, the prognostic significance of the intranodal lymphatic sinuses when measured in their totality in involved LNs, and the association of alterations in subcapsular sinus areas with superior distant metastasis-free survival in cancer-free and involved LNs in TNBC N+ patients. In summary, smuLymphNet presents a robust multiscale deep learning framework to automatically detect, localise and quantify histopathological immune features in WSI of LNs. By applying smuLymphNet to LNs of TNBC patients from clinical trials, and thereby further evaluating its clinical utility, smuLymphNet could be implemented into the diagnostic digital pathology workflow and, as such, aid in informing on a patient’s disease trajectory. Citation Format: Gregory Verghese, Mengyuan Li, Fangfang Liu, Amit Lohan, Nikhil Cherian, Patrycja Gazinska, Aekta Shah, Aasiyah Oozeer, Cheryl Gillett, Elena Alberts, Thomas Hardiman, Roberto Salgado, Samantha Jones, Louise Jones, Selvam Thavaraj, Sarah E. Pinder, Swapni Rane, Amit Sethi, Anita Grigoriadis. Multiscale Deep Learning framework to capture systemic immune features in lymph nodes predictive of triple negative breast cancer outcome [abstract]. In: Proceedings of the 2022 San Antonio Breast Cancer Symposium; 2022 Dec 6-10; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2023;83(5 Suppl):Abstract nr P5-01-01.