Multi-task learning to improve performance consistency in mammogram classification

Mickael Tardy, Diana Mateus

Keywords: Breast cancer, classification, multi-task learning, mammography

TL;DR: Multi-task learning including classification, reconstruction, segmentation, and regression objectives, allows for more consistent performances

Abstract: Breast cancer is the most prevalent cancer amongst women. Its regular screening, often based on mammograms, significantly reduces the mortality. Deep learning has shown good performances in coping with screening-generated imaging data, however there are still open questions related to the imbalance, noisiness, and heterogeneity of the data. We propose to address these challenges with Multi-Task Learning, combining tasks such as classification, regression, segmentation, and reconstruction. Our approach allows to obtain consistent performances of AUC $\approx 0.80$ across different vendors (including those unknown during training) on the primary breast cancer classification task, while fulfilling well secondary tasks including an $F_1$ score of $0.96$ on 4-class vendor classification, and $F_1$ score of 0.64 on 4-class density classification.

Registration: I acknowledge that acceptance of this work at MIDL requires at least one of the authors to register and present the work during the conference.

Authorship: I confirm that I am the author of this work and that it has not been submitted to another publication before.

Paper Type: recently published or submitted journal contributions

Primary Subject Area: Application: Radiology

Secondary Subject Area: Learning with Noisy Labels and Limited Data

Confidentiality And Author Instructions: I read the call for papers and author instructions. I acknowledge that exceeding the page limit and/or altering the latex template can result in desk rejection.