2020

"Understanding ROC (Receiver Operating Characteristic) Curve | What is ROC?"

"从价格双轨制，推演公立的衰落"

"沈向洋博士：三十年科研路，我踩过的那些坑"

"Making Your Neural Network Say “I Don’t Know” — Bayesian NNs using Pyro and PyTorch"

"Computer vision news December 2020"

AI applications in ultrasound: Select clips with good quality; Select only the important clips; Sorting clips according to what view and cross-ectionof the heart they show; detect features on the clips; taking measures; predict certain pathnologies;

Before training a deep learning model, it is vital to understand your data. Is it sufficiently variable to cover the application being developed? There are so many factors that come into play here, including the quality of the ultrasound, and characteristics of the patient, such as age, gender, and BMI, which can all effect the size of the organ being imaged. Once the data is ready, you need to work with your echo specialist or radiologist on the best labeling and annotation procedure. It is an iterative process which is partly about the algorithms and partly about the data. Try one approach, train a model, and then give the data back to the echo specialist who will look again and may make further suggestions.

Data is a key requirement for deep learning algorithms to work, but for ultrasound, data availability can be limited. Also, most applications involve heavy human interaction, whereas in computer vision, you work with a picture or video. “I personally see there’s a traditional imageguided interventional subfield and a robotic field,” Yipeng explains. “The robotic field is assuming robots will be controlling all medical devices in the future. With that as our end goal, we try to make our algorithms as automated as possible. We’re still very close, but these two fields will merge at some point.”