A new study published in Science Translational Medicine shows the Mirai mammography-based deep learning (DL) risk model is better at identifying high-risk cancer than previous models, reports HealthDay News.
Adam Yala, from the Massachusetts Institute of Technology (MIT) in Cambridge, and colleagues developed Mirai, a mammography-based DL model designed to predict risk at multiple time points and produced predictions that are consistent across mammography machines. Mirai was trained on a large dataset from Massachusetts General Hospital (MGH) and was tested on test sets from MGH, Karolinska University Hospital in Stockholm, and Chang Gung Memorial Hospital (CGMH) in Taoyuan City, Taiwan.
“Mirai demonstrated improved discriminatory capacity over the state-of-the-art clinically adopted Tyrer-Cuzick and prior deep learning approaches Hybrid DL and Image-Only DL,” the authors write.
Read more at HealthDay News.
Featured image: Fig. 1 Schematic description of Mirai. The four standard views of an individual mammogram were fed into Mirai. The image encoder mapped each view to a vector, and the image aggregator combined the four view vectors into a single vector for the mammogram. In this work, we used a single shared ResNet-18 as an image encoder, and a transformer as our image aggregator. The risk factor predictor module predicted all the risk factors used in the Tyrer-Cuzick model, including age, detailed family history, and hormonal factors, from the mammogram vector. The additive hazard layer combined information from both the image aggregator and risk factors (predicted or given) to predict coherent risk assessments across 5 years (Yr). (Courtesy: Science Translational Medicine)