Reproduce “Multi-task fusion for improving mammography screening data classification” on OpenResearch

About this paper

Maria Wimmer, Gert Sluiter, David Major, Dimitrios Lenis, Astrid Berg, Theresa Neubauer, Katja Bühler

A multi-task fusion pipeline was developed for mammography screening, integrating predictions and high-level features from diverse deep learning models specializing in breast density, findings, and lesion localization. This approach achieved improved patient-level classification for lesion presence and malignancy, with AUC increases of up to 0.04 and 0.029 respectively over naive baselines, while also providing intermediate, interpretable results.

Key insights

• Fusing information from distinct, task-specific models (e.g., density, general findings, lesion-specific) leads to superior patient-level predictions compared to single-task models or naive ensembling.
• High-dimensional feature-level fusion generally yields better performance than fusing prediction scores from individual models.
• Pre-training strategies, such as patch-wise pre-training for findings or view-wise pre-training for density, significantly improve the performance of individual task-specific models.
• Including breast density information within the fusion pipeline contributes to enhanced classification performance for both lesion and malignancy prediction.

Summary by alphaXiv.