Lesion Detection of Cerebral Infarction by a Machine Learning Model that Learned ‘Normal’ [Presidential Award Proceedings]
Akihiko Wada, Yuya Saito, Shinpei Kato, Akifumi Hagiwara, Shohei Fujita, Kotaro FUJIMOTO, Yutaka Ikenouchi, Kanako Sato, Toshiaki Akashi, Maki Amano, Koji Kamagata, Kanako K. Kumamaru, Atsushi Nakanishi, Michimasa Suzuki, Masaaki Hori, Shigeki Aoki · Japanese Journal of Magnetic Resonance in Medicine · 2020
We attempted to detect and diagnose cerebral infarction on diffusion weighted image (DWI) using a machine learning model trained using normal brain data. Our machine learning model consisted of two parts. One consisted of an autoencoder (AE), which learned only normal DWI information. This AE model did not reproduce the lesion ; the input image was subtracted and the generated image contributed to abnormal signal detection. The other part was a classifier constructed with SqueezeNet, which distinguished between infarct and normal areas using the original DWI image, AE generated images, and subtraction images. The accuracy of the combination of AE and SqueezeNet in 2-class classification for abnormal detection on DWI was 0.97.