MULTI-VIEW DEEP EVIDENTIAL FUSION NEURAL NETWORK FOR ASSESSMENT OF SCREENING MAMMOGRAMS

Naga Raju Gudhe, Mazen Sudah, Arto Mannermaa, Veli-Matti Kosma, Hamid Behravan

Keywords: Multi view fusion, Mammograms, Evidential learning, Deep learning

Abstract:

Mammography is an X-ray-based imaging technique widely used for breast cancer screening and early-risk assessment. A large number of mammograms are acquired in regular breast cancer screening programs. The assessment of mammograms is a tedious task and may be difficult to accomplish due to a shortage of expert radiologists in breast imaging. Artificial intelligence-powered algorithms, especially deep learning, could assist radiologists by automating the assessment, however, substantial trust needs to be established in incorporating such algorithms in real-world settings. The evidential neural networks algorithm provides an interpretable approach using Dempster-Shafter evidential theory that supports network predictive confidence. Recent studies have suggested that multi-view analysis improves the assessment of mammograms. In this study, we advance the multi-view assessment of mammograms by using a deep evidential neural network to address the following questions:

1. What is the effect of various pre-trained convolutional neural networks in extracting features from mammograms?
2. Which fusion strategies work better for the multi-view assessment of mammograms using a deep evidential learning framework?

The multi-view deep evidential neural network extracts features from each mammogram’s view using a pre-trained convolutional neural network. The extracted features are combined using Dempster-Shafer evidence theory for the following two classification tasks, mammogram density assessment in BI-RADS categories and mammogram finding as benign or malignant. We conducted extensive experiments using two open-sourced digital mammogram datasets, VinDr-mammo, and mini-DDSM, with 4,977 and 1,885 patients, each with four mammogram views, respectively. The results suggest that the multi-view approach outperforms the single-view by relative improvements of 2.99% and 2.64% for VinDr-mammo, and 6.51% and 8.75% for mini-DDSM datasets, in terms of F1-score, in mammogram density assessment and BI-RADS findings benign/malignant classification tasks, respectively. Our results show that the multi-view assessment of mammograms using a deep evidential fusion approach not only provides superior performance than the single-view assessment but also enhances trust in incorporating artificial intelligence-powered algorithms for the assessment of screening mammograms.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )