Bag of deep features for preoperative prediction of sentinel lymph node metastasis in breast cancer

Abstract

Breast cancer is the most common female malignancy among women. Sentinel lymph node (SLN) status is a crucial prognostic factor for breast cancer. In this paper, we propose an integrated scheme of deep learning and bag-of-features (BOF) model for preoperative prediction of SLN metastasis. Specifically, convolution neural networks (CNNs) are used to extract deep features from the three 2D representative orthogonal views of a segmented 3D volume of interest. Then, we use a BOF model to furtherly encode the all deep features, which makes features more compact and products high-dimension sparse representation. In particular, a kernel fusion method that assembles all features is proposed to build a discriminative support vector machine (SVM) classifier. The bag of deep feature model is evaluated using the diffusion-weighted magnetic resonance imaging (DWI) database of 172 patients, including 74 SLN and 98 non-SLN. The results show that the proposed method achieves area under the curve (AUC) as high as 0.852 (95% confidence interval (CI): 0.716–0.988) at test set. The results demonstrate that the proposed model can potentially provide a noninvasive approach for automatically predicting prediction of SLN metastasis in patients with breast cancer.