9P - Development and validation of a pathogenomics model to improve the risk stratification of breast cancer: A deep learning study

Background

The risk evaluation of Disease-free survival (DFS) is peculiarly challenging in breast cancer. In this study, we use the combination data of digital pathology, survival-related gene expression, and deep learning (DL) for the ranked prediction risk of DFS in breast cancer. In this process, in order to better predict the risk of disease-free survival after breast cancer, we combined DeepSurv model, ResNet50 model, and Attention-gate mechanism to propose a new deep learning model ROAD according to the structural characteristics of the data.

Methods

In this study, 925 invasive breast cancer patients from TCGA were included and were divided into the training cohort (n=741) and the validation cohort (n=184). This study included three phases to develop the DL model. In phase 1, according to the random forest algorithm, a high survival-related gene will be chosen, and then the DeepSurv model was constructed with deep learning survival-related gene features for DFS prediction. In phase 2, digital pathology features can be extracted by pre-trained ResNet50, and then the attention-gate is used to screen the features for DFS prediction. We design specific access from the above DL network by adding the gene information to the end of the attention-gate, which can be seen as the input of the DeepSurv, that can combine pathology with genomics to predict Disease-free survival risk degree after breast cancer surgery was further analyzed in phase 3.

Results

The deep learning model was significantly associated with DFS (P < 0.001), and it achieved C-index of 0.968(±0.003) on the training dataset. And the model showed significant improvement the C-index from 0.891 to 0.965(±0.04) for DFS in the testing cohort by comparing the gene expression matrix and pathology features with the combination.

Conclusions

The ROAD model can achieve an accurate risk evaluation of breast cancer patients and can be conveniently used to guide treatment decisions and predict DFS in breast cancer.

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.