Why is it hard to spot problems in the fetal brain?
Congenital malformations of the central nervous system, meaning the brain and spinal cord, are among the most common major malformations present at birth. During pregnancy scans, professionals called sonologists must spot them on grey-and-black images that take a lot of experience to interpret.
Deep-learning systems (AI that learns to recognise patterns from examples) are making progress in prenatal diagnosis. But their real effect on the work of sonologists had not yet been evaluated for the fetal brain.
How was the study run?
The team tested a system called PAICS in a three-way crossover randomised trial. In total, 709 fetal brain ultrasound images or videos were read by 36 sonologists with different levels of expertise.
Each reading was done in three ways. Unassisted: the professional reads alone. "Concurrent" mode: the AI is consulted at the very start of the assessment. "Second" mode: the AI is used only after a complete unaided interpretation, to double-check one's own answer.
What was observed?
The readers' average accuracy, 73% unassisted, rose to 80% in concurrent mode and 82% in second mode. Another measure, the AUC (a score from 0 to 1 where 1 means a perfect separation of normal and abnormal cases), went from 0.85 to 0.89 and then 0.90. These differences are statistically significant.
The flip side: the median reading time per case rose slightly in concurrent mode (from 6 to 7 seconds) and markedly in second mode (11 seconds), because you read everything twice. The authors conclude that the AI is more efficient in concurrent mode for all readers.
Key figures
- 73% → 82%average sonologist accuracy, unaided then with the AI after an unaided reading (80% with the AI from the start)
- 36sonologists who read 709 images or videos
- 6 s → 11 smedian reading time per case, unaided then in "second" mode (7 s in concurrent mode)
What does this change for patients and future parents?
Better detection could in principle allow abnormalities to be spotted earlier, to be confirmed and followed up. The tool acts like an automatic second opinion.
It is not an automatic diagnosis: a flagged abnormality must always be examined and confirmed by specialists.
What are the limits?
Readers looked at images and videos gathered for the study, not real live consultations. In real life, the large majority of scans are normal; here, abnormal cases were deliberately more common. The authors therefore expect that the AI might raise more false alarms in real practice, and that the accuracy gain could be different.
Because these malformations are rare, each type of abnormality is represented by only about 70 images or videos on average, which may not reflect the full variety or severity of lesions. The gains will need to be checked in wider, real-world conditions. The study measures reading quality, not consequences for pregnancies or children.
A pregnancy, some scans: Mima is with you
This AI acts as a second opinion for sonographers, helping them spot brain anomalies in the foetus. The medical team decides, and it is not available everywhere.
During a pregnancy, Mima can keep your scan appointments and the questions you want to ask, and bring them back at the right time. She also explains the words in the report. A question: "Does my sonographer use a tool to help read the images?"
Frequently asked questions
Does the AI replace the sonologist?
No. In the study the AI assists professionals, who remain the ones reading the images. A flagged abnormality must always be confirmed by specialists.
When is the best moment to use the AI?
Both modes improve accuracy, but the authors conclude that using it concurrently, from the start of the reading, is more efficient for all readers because it takes less time.
Is this tool used for my pregnancy?
The study does not say it is already routine. It tested the system on images with 36 sonologists. Ask your care team how your scan is interpreted.
What does AUC measure?
It is a score between 0 and 1 that measures the ability to tell normal from abnormal cases. It went from 0.85 unaided to 0.89 and 0.90 with the AI.
The original study
Lin M, Zhou Q, Lei T, et al. “Deep learning system improved detection efficacy of fetal intracranial malformations in a randomized controlled trial.” npj Digital Medicine, 13 October 2023. DOI : 10.1038/s41746-023-00932-6
Read the full study in npj Digital MedicineThis article summarises a scientific study and is not medical advice.