| Summary: | This thesis addresses the task of developing deep-learning-based tools for assessing placenta accreta spectrum disorders (PASD), a series of placental abnormalities that drastically affect maternal health, on placental ultrasound scans obtained from prenatal examinations.
The contribution of this thesis is a novel design of an image analysis pipeline consisting of computational modules that automatically search for clinically relevant sonographic patterns in placental ultrasound images and perform PASD assessment. This involves identifying informative image planes that clearly depict the anatomy around the placenta within a three-dimensional ultrasound volume; localizing the randomly distributed sonolucent spaces within the placenta (also known as placenta lacunae) caused by abnormal placental vasculature; detecting the utero-placenta interface where abnormal placenta adherence or invasion may occur; integrating ultrasound signs in a collaborative manner to classify PASD. This is the first work to attempt image-based automated prenatal assessment of PASD.
The computational modules are formulated as deep convolutional neural networks, whose parameters are gradually learnt by optimizing objective functions that minimize differences between network outputs and experts' reference annotations. Multi-scale representation learning is considered critical in localizing sonographic patterns with variable sizes and shapes such as patterns associated with placenta lacunae and utero-placental interface. The resulting deep learning models aggregate intermediate features from shallow to deep stages whilst exploiting global contextual cues to model placenta geometry. Apart from architectural refinement, methods are presented to further regularize network behaviours via auxiliary learning, based on additional supervision signals derived from experts' annotations. A complete, fully automated image-based PASD assessment pipeline is presented by effectively combining these computational modules to carry out plane-wise and volume-wise PASD assessment.
A cross-institutional placental ultrasound image analysis database is presented to validate the algorithms proposed in this thesis. The computational modules are compared against a number of competitive benchmark models. Assessment results are further compared to estimates made by a panel of experts who specialize in PASD diagnosis. The results suggest that the proposed PASD assessment pipeline has the potential to be integrated in clinical scenarios to assist in prenatal diagnosis.
|
|---|