The deep learning model estimates cancer risk from breast density

Breast cancer is the most common cancer affecting women worldwide. According to the American Cancer Society, about 1 in 8 women in the United States will develop breast cancer in their lifetime. While it’s impossible to completely prevent breast cancer, various medical organizations recommend regular screenings to detect and treat cases at an early stage. Breast density, defined as the proportion of fibro-glandular tissue in the breast, is often used to assess the risk of developing breast cancer. While various methods are available for estimating this size, studies have shown that subjective judgments made by radiologists based on visual analogue scales are more accurate than other methods.

Since expert evaluation of breast density plays an important role in breast cancer risk assessment, developing an image analysis framework that can automatically estimate this risk, with the same accuracy as experienced radiologists, is urgently needed. For this purpose, the researchers led by Prof. Susan M. Astley of the University of Manchester, UK, recently developed and tested a new deep learning-based model capable of estimating breast density with high precision. Their findings are published in Journal of Medical Imaging.

“The advantage of a deep learning-based approach is that it allows automatic feature extraction from the data itself,” explains Astley. “This is interesting for breast density estimation because we don’t fully understand why subjective expert judgment outperforms other methods.”

Typically, training deep learning models for medical image analysis is a challenging task due to limited data sets. However, the researchers managed to find a solution to this problem: instead of building a model from scratch, they used two independent deep learning models that were originally trained on ImageNet, a non-medical imaging dataset with over a million images. This approach, known as “transfer learning”, allowed them to train models more efficiently with less medical imaging data.

Using nearly 160,000 full-field digital mammogram images that were assigned a density score on a visual analogue scale by experts (radiologists, advanced practicing radiographers, and breast doctors) of 39,357 women, the researchers developed a procedure to estimate the density score for each mammogram image. . The goal is to take a mammogram image as input and generate a density score as output.

The procedure involves preprocessing the images to make the training process less computationally intensive, extracting features from the images processed with a deep learning model, mapping the features to a set of density scores, and then combining the scores using an ensemble approach to produce the final result. density estimate.

With this approach, the researchers developed a highly accurate model for estimating breast density and its correlation to cancer risk, while saving computation time and memory. “The model’s performance is comparable to that of human experts within limits of uncertainty,” said Astley. “In addition, it can be trained more quickly and on small datasets or subsets of large datasets.”

In particular, the deep transfer learning framework is useful not only for estimating breast cancer risk in the absence of a radiologist, but also for training other medical imaging models based on estimates of breast tissue density. This, in turn, could enable performance gains in tasks such as cancer risk prediction or image segmentation.

Read the Open Access article by S. Squires et al., “Automatic assessment of mammographic density using deep transfer learning methods,” J.Med. Imaging 10(2) 024502​​(2023), doi 10.1117/1.JMI.10.2.024502.