Automatic analysis of mammograms

We develop methods for image analysis of mammograms in close collaboration with the Cancer Registry of Norway and various breast diagnostic centres around the country.

We have extensive experience in this domain, and were one of the first institutes to apply deep learning techniques for this purpose. Our partnership with the Cancer Registry and regional diagnostic centres has enabled access to millions of mammograms labelled with confirmed cancer diagnoses, sourced from the national screening programme.

The image shows an x-ray of a breast in black, white and grey tones. The breast tissue is shown in white. A potential sign of cancer is located in a red box and further pinpointed with a red dot. — Caption: The image shows a breast where cancer has been detected with AI. The red square shows the region of the breast identified as suspicious, while the red marking indicates where cancer is likely to be located. Image: NR/The Cancer Registry of Norway.

Line Eikvil

Research Director BAMJO

Partner: The Cancer Registry of Norway

Further reading:

Visual Intelligence – visual-intelligence.no

BreastScreen Norway – kreftregisteret.no

The image shows an x-ray of a breast in black and grey to illustrate image quality analysis: Red dots form a line down the side of the pectoral muscle, the blue point indicates the bottom edge of the breast, and the yellow one shows the nipple. — Caption: A graph convolutional model has been used to detect important points in the image. The red points show the edge of the pectoral muscle, the blue point indicates the bottom edge of the breast, while the yellow one shows the nipple. All of these should be readily visible in the mammogram, correctly positioned within the image and in relation to each other. Image: NR/The Cancer Registry of Norway.

Explainable artificial intelligence (XAI)

Our models are at the forefront of breast cancer detection, surpassing academic and commerical counterparts. To ensure practical application and trustworthiness, it is important that humans are able to understand how the models make decisions. To that end, we use XAI. These techniques can be used to generate heatmaps that highlight which parts of the breast contribute to the model’s prediction. We also integrate XAI in the model itself in order to automatically identify the area of the breast that warrants an additional analysis in higher resolution by a second model.

Adaptable models for different datasets

A key feature of cancer is the uncontrolled growth and spread of abnormal cells. We are developing deep learning models that can extract information from earlier mammograms in order to detect changes that may be malignant. Another important issue is the models’ ability to generalise effectively across diverse datasets, including different populations, time periods and imaging equipment. Consequently, we focus on exploring methods that enable the models to adapt to various settings without needing to be retrained for each new domain.

Improving mammographic image quality

Our primary focus is cancer detection, but good image quality is a vital component in this process, and we are therefore developing methods to assist and improve analysis and image quality. Key factors to image quality are image blur, tissue density and breast positioning within the images. Presently, we are looking into the latter, breast positioning, and developing deep learning graph convolution models that can identify significant areas in the breasts. These models can aid radiographers in assessing mammogram quality, and are able to provide immediate feedback during imaging, for instance if key points are missing or a breast is positioned incorrectly.

Selected projects

The image shows stress fibres and microtubules in human breast cancer. Image by: Christina Stuelten, Carole Parent, 2011

Image analysis
Machine learning

Breast cancer detection with machine learning

The images shows the body of a woman in a pink shirt holding a bright pink ribbon which symbolises breast cancer awareness.

Image analysis
Machine learning