The desirable date of the birth of a child follows the full duration of pregnancy. According to WHO data, 15 million children are born prematurely every year, of which 1.1 million dies, unfortunately. In this talk, we will present how to improve prediction rate of the spontaneous preterm delivery using deep learning and computer vision methods.
Preterm delivery (preterm birth) is the most common cause of neonatal death. Despite great advancement in pregnancy care, frequency of preterm delivery does not decrease. Therefore, it is important to develop prediction methods that will assess risk of the preterm delivery and thus enable gynecologists to start appropriate prevention treatment.
Current diagnostic methods that assess risk of the spontaneous preterm delivery involve collection of maternal characteristics (via interview) and transvaginal ultrasound (US) conducted in the first and second trimester of pregnancy. Analysis of the US data is based on gynecologist's expertise, that is visual inspection of images, which is sometimes supported by hand-designed image features such as cervical length. Such approach is naturally prone to errors, thus approximately 30% of spontaneous preterm deliveries are not predicted. Moreover, 10% of predicted preterm deliveries are false-positives.
During the talk, we will present results of our project focused on improving prediction rate of spontaneous preterm delivery. To that end, we use a deep neural network architecture trained for segmenting prenatal ultrasound images and another network to classify the resulting cropped image. We validate our results on real-life cases as our research is done in a close collaboration with gynecologists from Warsaw University of Medicine - they provide us with constant feedback and assess the usability of our solution in a day-to-day gynecologists' routine.
Our talk is meant for intermediate machine learning researchers, engineers and students who are interested in the practical aspects of using deep learning for medical imaging.