“I know what normal looks like:” Microsoft trains AI to scan breasts

Microsoft’s research aims to bring higher accuracy to breast cancer screening using AI.

Breast cancer is the most common cancer among women globally. And in the US, about one in eight women will face a diagnosis in her lifetime.

Let’s be clear: AI will not replace doctors and radiologists anytime soon. However, it provides the opportunity to detect cancer earlier, reduce unnecessary interventions, and ultimately save more lives.

This week, Microsoft’s AI for Good Lab, in partnership with the University of Washington and Fred Hutchinson Cancer Center, published a study in Radiology that might actually signal a turning point in how AI gets used in breast cancer screening.

MRI screenings provide too many false-positive results

Magnetic Resonance Imaging (MRI) is one of the sharpest tools we’ve got for catching breast cancer early, especially for women with dense breast tissue. Paradoxically, tissue density raises the risk of cancer, but also makes it much harder to spot anything suspicious on standard mammograms.

However, MRI screenings often end with false positives, showing pathological lesions that are not necessarily cancerous. Most suspicious spots turn out to be benign, and that uncertainty leads to anxiety, additional scans, and often unnecessary biopsies.

Having that in mind, the researchers have created a technology called Fully Convolutional Data Description, or FCDD. This novel AI model focuses on what's typical instead of memorizing every permutation of cancer imagery.

Training on all possible cancer variants is a near-impossible task, especially in screening settings where cancer is rare. That’s why the model learns the patterns of normal breast tissue and flags outliers.

The study, which analyzed over 9,700 MRI exams, tested FCDD in screening-like conditions where only around 1.85% of scans contained actual cancer – a far cry from the curated datasets most AI models train on.

During the study, the model outperformed traditional approaches. It doubled the positive predictive value of standard AI models and reduced false positives by over 25%.

The technology also offers transparency on the vetting process. Instead of a black-box verdict, FCDD generates heatmaps that visually highlight areas of concern. These maps aligned with radiologists' retrospective annotations with 92% accuracy.

The model is not aimed at replacing radiologists

Microsoft and its research partners aren’t claiming that FCDD will replace radiologists. What they claim to be building is an assistive tool designed to triage, reduce workload, and make screening smarter.

“We are very optimistic about the potential of this new AI model,” said Savannah Partridge, a professor of radiology at the University of Washington and senior author of the study.

She emphasized that it works with both full diagnostic and abbreviated MRIs, meaning it could save time for both patients and providers.

The model is also open-source, which adds another layer of credibility. Researchers can explore the code, test it, and challenge it.