OP01.08: Exploring large prenatal image databases with a deep learning framework

Abstract

Image databases of large ultrasound centres often contain several million ultrasound images. Searching a large database for specific content may become impossible for human beings. To adress this issue we explored the potential of machine learning techniques. An open source machine learning library (PyTorch) was trained to recognise first trimester mid-sagittal views of the face in an completely unselected random sample of ultrasound images. For the training phase we selected 116 ideal positive, 133 weak positive and 956 negative images. For evaluation the developed algorithm was tested on 4908 randomly selected images from the entire database and compared with the evaluation by ultrasound experts. Judged by ultrasound experts, there were 46 positive images in 4908. 37/46 (80.4%) NT images were correctly recognised by the algorithm. 9/46 (19.6%) were not recognised. 96/4728 (2.03%) were detected false positive. From 96 false-positive images, 25 were clearly false-positive images (abdominal circumference, limbs, four-chamber view), that is 25/4732 (0.53%) of all negative images. The remaining false positive images were mid-sagittal views of fetal faces 15 to 23 weeks of pregnancy or mid sagittal views of the whole fetus. Despite relatively low training effort, the potential of modern image recognition algorithms in retrospective image reviews is impressive. We predict that deep learning frameworks can aid in investigating scientific issues in image databases of any imaginable size. Please note: The publisher is not responsible for the content or functionality of any supporting information supplied by the authors. Any queries (other than missing content) should be directed to the corresponding author for the article.

Ähnliche Arbeiten