Related Experiment Video
Updated: Jul 21, 2025
10:39
A Label-Free Segmentation Approach for Intravital Imaging of Mammary Tumor Microenvironment
Published on: May 24, 2022
2.4K
Leveraging Multi-Task Learning to Cope With Poor and Missing Labels of Mammograms
Mickael Tardy1,2, Diana Mateus1
1Ecole Centrale de Nantes, LS2N, UMR CNRS 6004, Nantes, France.
Frontiers in Radiology
|July 26, 2023
Summary
This study introduces a multi-task learning approach to improve breast cancer screening using mammograms. By weighting uncertain labels, the method enhances diagnostic accuracy for malignancy detection.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Oncology
Background:
- Accurate breast cancer screening relies on classifying mammograms as malignant or benign.
- Clinical data often lacks confirmed diagnoses, leading to label uncertainty and reduced training samples for Computer-Aided Diagnosis (CADx) systems.
- Label ambiguity in mammogram datasets hinders the development of robust CADx systems.
Purpose of the Study:
- To develop a deep-neural-network-based classifier that maximizes training samples by addressing label uncertainty in mammograms.
- To improve the performance of breast cancer screening through a multi-task learning framework.
- To incorporate class-specific uncertainty weighting during the training of deep neural networks for mammogram analysis.
Main Methods:
- A deep neural network was designed for multi-task learning, predicting malignancy, cancer probability, breast density, and image laterality in a single forward pass.
- A novel approach was introduced to weight training samples based on label uncertainty, preventing updates on ambiguous or missing data.
- The model was evaluated on public (INBreast) and private datasets, with further fine-tuning using data from the Susan G. Komen Tissue Bank.
Main Results:
- The multi-task learning approach demonstrated statistically significant improvements over baseline and state-of-the-art methods.
- The system achieved an Area Under the Curve (AUC) of 80.46% on a private dataset and 85.23% on the INBreast dataset for binary classification.
- Fine-tuning with clinical data further enhanced performance, showcasing the model's adaptability.
Conclusions:
- The proposed multi-task learning strategy effectively handles label uncertainty in mammogram datasets, increasing the number of usable training samples.
- This method offers a robust solution for developing more accurate CADx systems for breast cancer screening.
- The approach shows promise for improving diagnostic performance by leveraging multiple data sources and advanced deep learning techniques.
