Multi-Task Fusion for Improving Mammography Screening Data Classification.

Area of Science:

• Artificial Intelligence in Medicine
• Medical Imaging Analysis
• Deep Learning for Healthcare

Background:

• Machine learning and deep learning are crucial for computer-assisted medical prediction, particularly in mammography.
• Current methods often ensemble task-specific models for a comprehensive patient view.
• There is a need for advanced methods to integrate information from multiple mammography tasks.

Purpose of the Study:

• To propose and evaluate a novel pipeline approach for mammography analysis.
• To fuse predictions and features from task-specific deep learning models for enhanced patient-level prediction.
• To improve diagnostic accuracy in mammography by integrating diverse model outputs.

Main Methods:

• Developed a pipeline involving training individual, task-specific models.
• Implemented a multi-branch deep learning model to fuse predictions and high-level features.
• Utilized public mammography datasets (DDSM and CBIS-DDSM) for training and evaluation.

Main Results:

• Achieved an AUC of 0.962 for predicting any lesion and 0.791 for malignant lesions at the patient level.
• Fusion approaches demonstrated significant AUC improvements (up to 0.04) over standard model ensembling.
• The pipeline provides both global patient-level predictions and task-specific results related to radiological features.

Conclusions:

• The proposed pipeline effectively fuses information from multiple deep learning models for superior patient-level mammography prediction.
• This approach offers a significant advancement over traditional model ensembling in medical imaging.
• The system aims to closely support radiologists by integrating comprehensive predictions into their reading workflow.