Deep learning rivals fellowship-trained radiologists at segmenting breast cancers on MRI
A deep learning network trained on a large set of imaging exams performed as effectively as fellowship-trained radiologists when segmenting breast cancers on MRI scans.
Breast tumor segmentation offers additional insights on key features—shape, texture, etc.—that can greatly enhance a patient’s diagnosis and prognosis. Unfortunately, the task is time-consuming, which has prohibited radiologists from routinely assessing tumor volumes.
On Wednesday, researchers took a step toward overcoming these constraints, training algorithms on more than 60,000 individual breast scans. The top performing 3D U-Net platform equaled manual segmentations undertaken by a number of expert rads but did so automatically.
The results may have real-world implications for patients and the ever-busy imaging professional, the authors explained Dec. 15 in Radiology: Artificial Intelligence.
“With reliable, fully automated segmentation, the overall clinical workflow could be improved, and such segmentation could aid radiologists in tumor detection and diagnosis,” Lukas Hirsch and Yu Huang, both with the City College of the City University of New York, and co-authors added.
The group retrospectively collected 38,229 exams from 12,475 women treated at a single institution between 2002 and 2014. Across the 64,063 individual images, radiologists segmented 2,555 cancers and 60,000-plus benign tumors. An additional 250 cancers were independently segmented and included in a testing dataset.
The 3D neural network achieved a similar Dice score (0.77) compared to the fellowship radiologists, who achieved scores ranging from 0.69-0.85, the authors reported.
Hirsch, Huang and colleagues highlighted two distinct features of their U-Net study: it relied on a full 3D network that captures features missed during 2D processing and avoided sampling artifacts common in conventional architectures.
“In conclusion, when trained on a sufficiently large dataset, a 3D U-Net can segment breast cancers with performance comparable to fellowship-trained radiologists,” the researchers added. “The network produces detailed 3D segmentations in routine clinical MRI. The code and pretrained network have been made freely available.”
Matt joined Chicago’s TriMed team in 2018 covering all areas of health imaging after two years reporting on the hospital field. He holds a bachelor’s in English from UIC, and enjoys a good cup of coffee and an interesting documentary.