Breast cancer diagnosis from contrast-enhanced mammography using multi-feature fusion neural network

Abstract

Objectives: To develop an end-to-end deep neural network for the classification of contrast-enhanced mammography (CEM) images to facilitate breast cancer diagnosis in the clinic.

Methods: In this retrospective mono-centric study, patients who underwent CEM examinations from January 2019 to August 2021 were enrolled. A multi-feature fusion network combining low-energy (LE) and dual-energy subtracted (DES) images and dual view, as well as bilateral information, was trained and tested using a large CEM dataset with a diversity of breast tumors for breast lesion classification. Its generalization performance was further evaluated on two external datasets. Results were reported using AUC, accuracy, sensitivity, and specificity.

Results: A total of 2496 patients (mean age, 53 years ± 12 (standard deviation)) were included and divided into a training set (1718), a validation set (255), and a testing set (523). The proposed CEM-based multi-feature fusion network achieved the best diagnosis performance with an AUC of 0.96 (95% confidence interval (CI): 0.95, 0.97), compared with the no-fusion model, the left-right fusion model, and the multi-feature fusion network with only LE image inputs. Our models reached an AUC of 0.90 (95% CI: 0.85, 0.94) on a full-field digital mammograph (FFDM) external dataset (86 patients), and an AUC of 0.92 (95% CI: 0.89, 0.95) on a CEM external dataset (193 patients).

Conclusion: The developed multi-feature fusion neural network achieved high performance in CEM image classification and was able to facilitate CEM-based breast cancer diagnosis.

Clinical relevance statement: Compared with low-energy images, CEM images have greater sensitivity and similar specificity in malignant breast lesion detection. The multi-feature fusion neural network is a promising computer-aided diagnostic tool for the clinical diagnosis of breast cancer.

Key points: • Deep convolutional neural networks have the potential to facilitate contrast-enhanced mammography-based breast cancer diagnosis. • The multi-feature fusion neural network reaches high accuracies in the classification of contrast-enhanced mammography images. • The developed model is a promising diagnostic tool to facilitate clinical breast cancer diagnosis.

Keywords: Breast neoplasms; Contrast media; Deep learning; Mammography.