ISBI Daily - Thursday

Oral Presentation: Deep Learning with Ultrasound Physics for Fetal Skull Segmentation 8 Thursday ISBI DAILY Juan says that segmentation of fetal imaging can be challenging and so is not very common. Acquiring the images is complicated so there are not many volunteers. Also, 2D ultrasound is still the preferred imaging method for fetal screening, but it can be very subjective. For example, the sonographer needs to look for the specific plane of the head to extract manually those biometrics that will tell them that the fetus is growing properly. Using 3D ultrasound , the whole head can be acquired in a single shot. When segmenting the skull in a 2D image it is practically as easy as drawing a circle; however, the challenge in 3D segmentation is you have to go slice by slice. Juan explains: “ What we did here is try to automate this process using a state-of-the-art segmentation architecture: deep learning, of course! Nowadays, everything is deep learning. In particular, we are using a U-Net architecture – it was presented a couple of years ago and it’s proving to be very good in segmenting images .” The reality of medical imaging throws up more challenges and Juan says that ultrasound is one of the most challenging – images can suffer from speckle, shadow, fading, low contrast, and low signal-to-noise ratio. To combat this, he tells us they had to do some fine tuning to improve the audio level. They embedded in the architecture some information that they already knew about the physics of ultrasound image acquisition . Ultrasound is like a radar, which throws Juan Cerrolaza is a Marie Curie fellow at Imperial College, London. He speaks to us ahead of his presentation today on segmentation of the fetal scalp in 3D ultrasound, which is part of a bigger project on fetal imaging called iFind . Juan Cerrolaza “ …to automate this process using a state-of-the-art segmentation architecture: deep learning, of course! ”