Assessments of cancer-free lymph nodes for the prediction of disease progression

Abstract

Abstract Background Histological changes in the cancer-free lymph nodes (LNs) can add to the risk prediction of developing distant metastasis for LN-positive breast cancer patients. As Whole Slide Images (WSI) demonstrated suitability for detection of LN metastasis with high accuracy based on Convolutional Neural Networks (CNN), we extended the assessments to cancer-free LNs to determine whether early signs of disease progression could be identified. Methods Clinico-pathological data, including level of LNs in the axilla, were available for 143 LN-positive breast cancer patients. All resected LNs were digitised (×40 magnification), totalling to 1,054 WSI. CNN models were based on standard ResNet architecture, and trained using 224x224 sized WSI patches, to classify germinal centres (GC), follicles, sinus histiocytosis, and adipose tissue as defined by 2 independent pathologists. To increase the robustness of the CNN, training data was augmented using image rotation and colour perturbation. A weighted cross entropy was used as a training loss function to account for the imbalance in the four types of patches. Confusion matrix and one-vs-all AUC scores were computed on WSIs patches extracted from held-out test patients to assess the discriminative power of the trained CNN models. Multivariate and cross-validated L2-regularised proportional hazard regression analyses were used for outcome prediction. Results An increased number and smaller sized GCs in cancer-free LNs, the prevalence of metastatic LNs in the low axilla, the location of GCs in metastatic LNs, and lymphocytic infiltration at the tumour invasive front were all indicative of developing distant metastasis in LN-positive breast patients. Metastatic LNs were found exclusively in the low axilla in 77% of patients. Amongst the 913 H&E WSI of cancer-free LNs, 21% presented with GC. Patch level AUC scores of detecting each type of tissue region ranged between 0.82 to 0.94 on held-out patients quantifying the effectiveness of CNN to be used as topological feature detector. Conclusions Addition planned WSI analyses of large breast cancer cohorts will expand on these data, capturing histological features in LNs, with the aim of identifying high-risk and low-risk disease, and aiding in pathological prognostication. Legal entity responsible for the study The authors. Funding Breast Cancer Research Trust, Breast Cancer Now, CRUK. Disclosure All authors have declared no conflicts of interest.