Tardy M, et al., 2021

Mammograms are the primary method used to screen for breast cancer, helping detect the disease early when it’s most treatable. As more people access screening, radiologists face a growing workload, which has led to increased interest in artificial intelligence (AI) tools to help them interpret images.

New technologies, like AI-based computer-aided diagnosis (CADx) systems, aim to estimate cancer risk or flag suspicious images. Training AI models require large sets of well-labeled mammograms so they can learn to make accurate predictions.

However, in real clinical settings, those clean labels are often missing. Some patients don’t have follow-up diagnoses, and reports can be inconsistent. This makes it hard to train effective AI systems.

To address this gap, the authors developed an approach that lets AI learn from mammograms even when labels are uncertain or incomplete. The method is called multi-task learning (MTL), and it allows the system to learn several related tasks at the same time instead of just focusing on cancer detection. This gives the AI system more data to learn from, improving its ability to detect cancer even when it doesn’t have perfect information.

Full-Text Journal Article PubMed

What did they study?

The researchers designed an AI system that learns from mammograms even when some labels are missing or unreliable. To do this, they trained an AI model to perform five tasks at once:

1. Decide whether the breast image is cancerous or not.
2. Estimate how likely the case is to be cancer (based on a six-level rating system used by radiologists).
3. Determine how dense the breast tissue is (which is a known risk factor for breast cancer).
4. Identify the angle from which the mammogram was taken (since images are taken from different angles).
5. Recreate the original image

• A private dataset of over 2,500 mammograms with full annotations.
• A public dataset called INBreast with 410 mammograms.
• A larger dataset from the Susan G. Komen Tissue Bank (KTB), containing images from healthy women. These KTB images were used to further improve the system’s performance.

What did they find?

1. Their AI system outperformed other state-of-the-art models. When tested on two different datasets, it achieved higher accuracy for cancer detection than other models, even those trained on much larger datasets.
2. Including multiple tasks significantly improved performance. When the system was trained to predict only whether a case was cancerous, it performed worse than when it also predicted other factors like breast density or view angle.
3. Their method held up even when trained with fewer cancer-labeled examples. This suggests it can still learn effectively in real-world situations where many images have missing or unclear information.
4. The system was able to identify non-cancerous cases with a low false-negative rate. This could help reduce unnecessary follow-ups or biopsies in patients without cancer, saving time and resources in clinical settings.

Why is this study important?

The researchers also improved something called explainability—meaning the system’s internal reasoning becomes easier to understand. For example, images with similar characteristics (like dense tissue or cancer risk) are grouped together in the system’s analysis, helping researchers and doctors see how it’s making decisions.

More broadly, this approach offers a pathway to building smarter, more flexible AI tools that can adapt to the realities of clinical data and can improve early detection in patients.

Full citation

Tardy M,Mateus D Leveraging Multi-Task Learning to Cope With Poor and Missing Labels of Mammograms.. Frontiers in radiology 2021; 1. PMID:37492176.