← All decks
The first 25 slides, exactly as they appear. The full deck has 99 content slides.
Radiology
Artificial Intelligence Approaches in Obstetric Imaging
Built from Obstetric Imaging
What’s inside
8 sections · 99 slides
Overview
- How this topic is organised
Why computers are being brought into the pregnancy scan
The problems in obstetrics, and what a learning machine can add to routine imaging
- Problems facing obstetric care
- What a learning machine adds
- Ultrasound as the mainstay of obstetric imaging
- Routine scans as an untapped data resource
- Three stated goals for artificial intelligence in pregnancy care
- Scope of this account
The vocabulary of artificial intelligence
A clinician-friendly guide to the terms you will meet in every AI paper
- Why clinicians need the terminology
- Artificial intelligence, machine learning and neural networks as nested fields
- Artificial intelligence defined
- Algorithms as input-to-output machines
- Machine learning defined
- Machine learning applied to clinical data
- Deep learning defined
- How the terms nest
- Training, validation and testing
- Testing and out-of-distribution data
- The training sequence
- Convolutional neural networks
- Convolutional neural network architecture
- Reading the network diagram
- From input image to a scored classification
- U-Net in biomedical image segmentation
- Two families of deep learning
- Supervised learning defined
- Supervised learning: labelling and training
- Supervised learning: validation and testing
- Advantages of supervised learning
- Disadvantages of supervised learning
- Unsupervised learning defined
- Unsupervised learning in practice
- Advantages of unsupervised learning
- Disadvantages of unsupervised learning
- Bias persists even without human labels
- Patient privacy and locked algorithms
- Supervised versus unsupervised learning
Computer vision and its clinical tasks
The five things a machine can do once it can look at a picture
- Computer vision defined
- Where computer vision is already used
- Five computer vision tasks in obstetric imaging
- Classification
- Segmentation
- Pixel-level labelling of an image
- What the segmented image demonstrates
- Segmentation applied to obstetric ultrasound
- Diagnosis
- The five fetal brain abnormalities in that model
- Quality assessment
- Guidance
- Guidance and the detection of cardiac abnormalities
- Safeguards on guidance features
What makes obstetric ultrasound a special case
Why the imaging test that suits pregnancy best is the one that suits a computer least
- How radiology adopted artificial intelligence
- Why computed tomography, magnetic resonance and x-ray suit AI
- Operator dependence in ultrasound
- Obstetric ultrasound in practice
- Extra obstacles specific to obstetric scanning
- From handheld acquisition to the need for guidance
- Guiding the operator during the scan
- Multiple frames from a single obstetric ultrasound examination
- What the frames show about consistency
- Obstetric imaging as a big-data resource
- Longitudinal imaging within one pregnancy
- The stated goal for clinical AI in ultrasound
Where the research stands today
What has actually been built and published in obstetric ultrasound
- Maturity of the evidence
- What most obstetric AI studies set out to do
- Published work on fetal biometry
- Published work on fetal neurosonography
- Published work on fetal cardiac imaging
- Placental texture analysis and hypertensive disorders
- 76.6%
- Reading those two numbers
- Three goals for AI-assisted image quality
- Scanning without a trained sonographer
- The purpose of that low-resource protocol
- Home monitoring and telehealth
Opportunities that remain open
Maternal anatomy, preterm birth, and the data types nobody has combined yet
- Maternal anatomy and preterm birth
- Artificial intelligence as a hypothesis generator
- Cervical texture analysis in twin pregnancy
- Supporting studies on cervical texture
- Measuring the cervix on transvaginal ultrasound
- Cervical length and the anterior uterocervical angle
- The classical methods used in that model
- What the two-marker model showed
- A more detailed cervical segmentation scheme
- Anatomy labelled on a transvaginal scan in pregnancy
- Why the bladder is labelled alongside the cervix
- Predicting preterm birth from cervical shape
- Data types still to be brought together
- Three-dimensional and quantitative ultrasound
- Reducing inequity in scan quality
- Obstetrics catching up with radiology
- Collaboration between clinicians and engineers
- What adoption will mean for practice
Summary
The points worth carrying out of this topic
- Key points: the vocabulary
- Key points: the clinical tasks
- Key points: the obstetric problem and its promise
- References
- References (continued)
- References (continued)
- Obstetric Imaging: Fetal Diagnosis and Care, 2nd Edition