By: Kimberly Rodrigues
A team of researchers from the University of Manchester has developed a new deep-learning model for estimating breast density, which may aid in predicting the risk of breast cancer.
Their approach uses automatic feature extraction from training data and transfer learning, making it well-suited for breast density estimations.
Breast density refers to the proportion of fibro-glandular tissue within the breast, and is a key factor in assessing breast cancer risk.
The team published their findings in the Journal of Medical Imaging, detailing the use of two deep learning models initially trained on a non-medical imaging dataset, which were then trained on their own medical imaging data.
Limited datasets make it difficult to train and build deep learning models from scratch. To tackle this, experts, including radiologists, advanced practitioner radiographers, and breast physicians, used visual analogue scales to assign density values to 160,000 mammogram images from 39,357 women.
The researchers developed a procedure that could estimate a density score by feeding a mammogram image as an input, using pre-processing techniques to make the training process less computationally intensive. They then used an ensemble approach to combine the scores and produce a final density estimate.
The researchers developed highly accurate models that estimate breast density and its correlation with cancer risk while conserving computation time and memory.
“The model’s performance is comparable to those of human experts within the bounds of uncertainty,” said lead researcher Susan M. Astley. “Moreover, it can be trained much faster and on small datasets or subsets of the large dataset.”
The researchers emphasised that their framework is not only applicable for estimating breast density and cancer risk, but can also be used as a basis for training other medical imaging models.
Breast cancer is the leading cancer among women across the globe, and while there are different ways to estimate breast density, studies have shown that subjective assessments by radiologists using visual analogue scales are the most precise method available.
(With inputs from PTI)
