Research press release


Scientific Reports

Cancer: State-of-the-art AI could improve breast cancer detection



今回、Dezso Ribliたちの研究グループは、最先端の深層学習技術を用いた、改良型コンピューター支援検出(CAD)システムを提案している。この改良型CADシステムは、トレーニングを行うことで乳房病変を検出し、その位置を特定できる。CADシステムは過去20年間にわたり、乳房X線像の解析と、検討を要するような疑わしい領域の標識によって放射線科医による乳がんの検出に役立ってきた。ところが、こうした技術は、高い利用コスト(米国で年間4億ドル)を伴い、その有用性については議論がある。



Deep learning artificial intelligence that is widely used to detect objects in images could improve early breast cancer detection, according to a study inScientific Reports.

Deep learning is a subset of machine learning. Its networks are inspired by knowledge of how biological brains, such as those of humans or animals, work. Deep learning networks ‘learn’ from datasets annotated by humans to reach image recognition capabilities similar to those of humans.

Dezso Ribli and colleagues propose an improved Computer Assisted Detection (CAD) system based on state-of-the-art deep learning that can be trained to detect and localize breast lesions. CAD systems have been used over the past two decades to help radiologists detect breast cancer by analyzing mammograms and marking suspicious regions which are then reviewed by a radiologist. However, use of these technologies is expensive ($400 million a year in the US) and their benefits remain controversial.

When tested on a dataset of 115 mammography cases (2 of 4 images per case) with proven cancers, the authors’ improved CAD system classified 90% of malign lesions with very few false positives and without human intervention. Current screening methods, which include assessment by radiologists, correctly detect 77-87% of cancers.

The findings suggest that expensive, traditional CAD methods may be replaced by less expensive deep learning methods that are currently being used to recognize objects - such as dogs and cars - in traditional images, to help radiologists detect more cancers. However, the authors caution that so far, they were only able to test their method on a small dataset of images with proven cancers.

doi: 10.1038/s41598-018-22437-z


メールマガジンリストの「Nature 関連誌今週のハイライト」にチェックをいれていただきますと、毎週各ジャーナルからの最新の「注目のハイライト」をまとめて皆様にお届けいたします。