7P - Predicting lymph node metastasis and molecular subtype from pathology united with genomics in breast cancer with multitask deep learning

Background

In early-stage of breast cancer, the analysis of lymph node metastasis(LNM) and molecular subtypes is conductive to clinical diagnosis and treatment decision-making. Here we proposed a new Deep learning network architecture, called DeepCALM, Conducted with Attention mechanism to united genomics and pathology, that can simultaneously predict LNM and Molecular subtypes.

Methods

In this retrospective study, DeepCALM was designed to analyse the breast cancer subtype and the degree of metastasis, which include the feature extraction from the digital pathological whold-slide images(WSIs), feature fusion between the feature of WSIs and RNA expression matrix, and muti-task classification. The WSIs and RNA expression data are all from The Cancer Genome Atlas(TCGA), and involved 916 breast cancer patients (2536 WSIs) that were split into training cohort of 734 patients (2021 WSIs) and validation cohort of 182 patients (515 WSIs), and the 219 survival-related gene that screened with the Random forest algorithm from training cohort. Model performances were evaluated using area under the curve (AUC), sensitivity, and specificity.

Results

DeepCALM achieved favourable accuracy for the classification of LNM(macro-average AUC of 0.977, sensitivity of 0.647, specificity of 0.953) and molecular subtypes(macro-average AUC of 0.960, sensitivity of 0.628, specificity of 0.941) in the training cohort, and LNM(AUC of 0.986, sensitivity of 0.647, specificity of 0.953) and molecular subtypes(AUC of 0.955, sensitivity of 0.628, specificity of 0.941) in the validation cohort (all p < 0.001).

Conclusions

The difference between LNM and molecular subtypes of breast cancer can be precisely distinguished by DeepCALM, which helps clinicians effectively determine the stage of tumor diffusion and speed up the diagnosis and accuracy, and then offer treatment or surgery plans.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

The authors.

Funding

Has not received any funding.

Disclosure

All authors have declared no conflicts of interest.