Revenue Cycle Insider

Discover how AI is trained to read mammograms like a radiologist.

The mammogram screening process captures X-ray pictures of the breasts. In a standard 2D mammogram, doctors take two pictures of each breast from different angles: the craniocaudal (CC) view (top-to-bottom) and the mediolateral oblique (MLO) view (side angle).

Radiologists use this dual-view technique to compare both images of each breast to detect abnormalities. Radiologists typically interpret four mammogram images per patient. In 2019 alone, 27.3 million mammograms were ordered or provided at office visits according to the 2019 National Ambulatory Medical Care Survey.

This is where new artificial intelligence (AI) tools can help doctors detect breast cancer more accurately. Read this first part of a two-part article series to learn how researchers are teaching AI to compare both views of each breast just like radiologists.

Learn if LLMs Can Read Mammograms

Tools like ChatGPT or Claude are a type of AI called large language models (LLMs). But if you tried uploading a mammogram to ChatGPT, it can’t generate a reliable diagnosis because the model (the “brain” of the AI) is not trained to perform that task. AI mimics the way humans learn and recognize patterns, so the AI must be trained first to recognize the patterns of breast cancer.

Teaching AI to analyze two views of each breast together can be especially difficult because breast compression during a mammogram deforms tissues, making it a challenge to digitally align the two views. A 2025 study by Pelluet et al. addressed this exact challenge by training AI to analyze both views without requiring precise image alignment.

Step 1: Gather the Training Data

Training AI to recognize breast cancer takes a lot of data. In the Pelluet study, researchers used thousands of mammogram images from both private and public datasets. The images came from mammography machines made by different manufacturers to ensure the AI model could work with images produced by different equipment.

Step 2: Label What’s Cancer and What’s Not

The researchers used a mix of normal, benign, and malignant cases. Different types of lesions were marked, as well as labels of “Cancer” and “No Cancer.” This tells the AI to associate certain visual patterns in a mammogram with cancer. Researchers used the breast imaging reporting and data system (BI-RADS) scoring system. The images labeled “Cancer” indicated a BI-RADS score of greater than 3 and “No Cancer” indicated a BI-RADS score of less than or equal to 3.

Step 3: Examine the Train-Test Split

When training AI, researchers can’t use all the data to teach the AI without the risk of “overfitting.” Overfitted models perform well only on the data on which they were trained. Instead, researchers use a portion of the images that the AI has never seen before to test whether the AI model really learned how to predict cancer or just memorized the training images. In the Pelluet study, the AI learned from 84 percent of the training images and was tested on the remaining 16 percent.

Step 4: Put In the Work

The AI trained for 15,000 cycles (called “epochs”), learning to compare both mammogram views of each breast. The AI divided each mammogram into small regions, then drew connections between potentially matching areas in the two different views. The AI looked at both global and local structures, which means it learned to look at an image in a specific area (like zooming in on a tumor) and also assessed the breast image as a whole.

The compression of breasts during a mammogram makes digital alignment of the two images being compared difficult during two-view analysis. Researchers addressed the issue by linking every suspicious region in one image to every region in the other image, sidestepping the need for digital alignment of the two images. To manage the massive number of resulting connections from high-resolution images, they used a technique called sparse bipartite graph attention network (SBIP), which ensured the AI focused on the connections relevant to detecting cancer while ignoring the irrelevant ones.

Grading: Since most mammograms are normal, researchers face an issue. In some datasets, as few as one in 20 images show cancer. This imbalance means the AI could achieve high accuracy scores by outputting “no cancer” every time. To solve this, the researchers used a technique called weighted cross-entropy loss.

The technique works like a grading system where the mistakes have different penalty weights. When the AI misses cancer (a false negative), it receives a severe penalty. In contrast, a false alarm on a normal mammogram (a false positive) results in a smaller penalty. The AI learns by avoiding penalties during training, thereby producing predictive models that can accurately detect cancer in mammograms.

Next month: Revenue Cycle Insider examines the results of the Pelluet study and looks at AI’s role in the future of healthcare.

Angela Halasey, BS, CPC, CCS, Contributing Writer