Predicting premature birth from MRI using deep learning

博士项目的目标：从胎儿核磁共振和出生指标预测妊娠和出生时体重的深度生成模型重点是使用模型解释来提高对早产和晚期胎儿生长限制因素的理解链接预测算法和获取以通知未来方案的发展，从而促进临床过渡。项目描述：人类怀孕和分娩是生命中最迷人的过程之一，但也给母亲和未出生的孩子带来重大风险。早产和胎儿生长受限的后果是严重的和终生的[1]。人类发育在很大程度上是隐藏的：临床上可用的超声波筛查只能瞥见子宫。对合适的妊娠何时开始分娩、胎儿的大小以及最有可能发生的分娩类型的先进知识，将允许在充分准备和监测所有孕妇(尤其是高危孕妇)方面逐步改变。胎儿磁共振技术的最新进展[2，3]已经能够成功地从结构和功能上评估胎儿器官和胎盘。这些结果显示胎儿生长受限时T2*值改变，早产威胁下妊娠时肺体积减小。动态信息，如胎儿运动和肺呼吸练习的频率，是早产的一个指标。胎儿核磁共振扫描包含大量信息，通常在多个平面上覆盖子宫。然而，目前只分析了所获得的数据的子集，例如只分析了大脑结构或胎盘附着。这在一定程度上是因为缺乏适当的分析工具，能够从获得的全部信息中受益。然而，无论扫描指示如何，婴儿将于何时出生以及体重将有多重的问题与所有分娩的临床计划高度相关。对这些问题的预测将允许优化产前护理和分娩时间点。最近的深度学习(DL)[4]技术提供了两个重要的机会：它们允许在不进行简化假设的情况下从复杂的图像数据集中寻找模式，同时可以调整以从图像返回显著的特征来回答这些问题。因此，该项目的目标是使用生成性深度学习来开发鸭嘴兽(通过妊娠全子宫MRI扫描预测妊娠长度、大约体重和分娩类型)：一种询问关于每次怀孕、分娩方式和时间的基本问题的工具。一个重要的重点将是利用模型可解释性的最新发展，为改进的获取策略提供信息，这些策略将持续绘制出生前最后几周的发展图。King‘s大型胎儿项目提供的广泛数据库提供了具有良好特征的训练数据，这对深度学习至关重要。

Aim of the PhD Project:Deep Generative Modelling for prediction of gestation and weight at birth from fetal MRI and birth metricsEmphasis on the use of model interpretation to improve understanding of elements involved in preterm birth and late fetal growth restrictionLink prediction algorithm and acquisition to inform future protocol developments and thus facilitate clinical transition.Project Description:Human pregnancy and birth are among the most fascinating processes in life but also carry significant risks for both mother and unborn child. Consequences from preterm birth and fetal growth restriction are severe and lifelong [1].Human development occurs largely hidden: clinically available Ultrasound screening only catch glimpses into the womb. Advanced knowledge of the appropriate gestation when the birth should start, the size of the fetus, as well as the type of delivery most likely to occur, would allow for a step change in adequate preparation and monitoring of all but especially high-risk pregnancies.Recent advances in fetal MRI techniques [2,3] have successfully been able to assess the fetal organs and placenta both structurally as well as functionally. These show altered T2* values in pregnancies with fetal growth restriction and reduced lung volume in pregnancies threatened by preterm labour. Dynamic information such as the frequency of fetal motion and lung breathing exercises is an indicator for preterm birth. Fetal MRI scans contain a vast amount of information and typically cover the uterus in multiple planes. However, only subsets of the data acquired are currently analysed, e.g. only the brain structure or placental attachment. This is in parts due to the lack of appropriate analysis tools able to benefit from the full extent of obtained information.However, regardless of scan indication the questions of when the baby will be born and how much they will weigh is of high relevance for clinical planning of all births. A prediction for these questions would allow to optimize antenatal care and time point of delivery.Recent deep learning (DL) [4] techniques provide two important opportunities: They allow to seek patterns from complex image data sets without making simplifying assumptions, and at the same time can be tuned to return features from the images which are salient to answering these questions. Thus, the goal of this project is to use generative Deep Learning to develop PLATYPUS (Prediction of Length of gestation, Approximate weight and TYpe of delivery from Pregnancy whole-Uterus MRI Scanning: a tool for asking fundamental questions about each pregnancy, mode and time of delivery. A significant emphasis will be to leverage recent developments in model interpretability to inform improved acquisition strategies which will continuously map development in the last weeks before birth.The extensive database available from large scale fetal projects at King's provides the well characterized training data which is critical for deep learning.