Healthcare · Medical imaging and neurology
Epilepsy lesions on MRI that almost nobody can see
Focal cortical dysplasias are one of the most common causes of drug-resistant epilepsy, and they are notoriously hard to spot on an MRI. Only a handful of specialists worldwide can reliably find them, so many patients with treatable epilepsy never get the right diagnosis. We built a 3D neural network that detects them with higher sensitivity than expert visual analysis.
Key results
Client: Evangelisches Krankenhaus Alsterdorf
The Hamburg Epilepsy Center at Evangelisches Krankenhaus Alsterdorf, a leading German institution for epilepsy diagnostics and treatment.
- Industry
- Healthcare and neurology
- Use case
- Epilepsy lesion detection on MRI
- AI approach
- 3D CNN with autoencoder regularisation
- Data
- MRI brain scans, the largest FCD training dataset in 2021
- Validation
- Prospective clinical validation, two peer-reviewed papers
- Learning
- Continual, without catastrophic forgetting
In short
A diagnosis that depended on a handful of people
- Focal cortical dysplasias cause drug-resistant epilepsy, and when they are found and surgically removed, many patients become seizure-free.
- Unlike a tumour they have no clear boundary. They vary in location, size and shape and blend into the surrounding tissue, so finding one takes rare specialist expertise.
- We built a 3D convolutional neural network that analyses the MRI in three dimensions, trained on the largest FCD training dataset at the time of its first publication in 2021.
- The current model reaches 90 percent sensitivity at 70 percent specificity, and both development stages are published in the peer-reviewed journal Epilepsy Research.
The starting point
The challenge
A critical diagnostic step, identifying FCDs on MRI, depended on a tiny pool of global specialists. Manual analysis was time-consuming, subjective and out of reach for most hospitals, and patients were being missed.
Focal cortical dysplasias are malformations in the cortex and one of the most common causes of drug-resistant epilepsy. Found and surgically removed, many patients become seizure-free. The difficulty is finding them.
A tumour usually shows up as a clearly visible mass. An FCD does not. They vary widely in location, size and shape, and they blend into the surrounding brain tissue without a clear boundary. Detecting one takes deep, specialised expertise in epileptology, and the number of people worldwide who can do it reliably is very small.
The result is that many patients with treatable epilepsy never receive the right diagnosis.
The build
What we built
Working directly with the neurologists and epileptologists at the Hamburg Epilepsy Center, we developed a system that detects and segments FCDs in 3D MRI brain scans automatically.
A neural network built for this one problem
We designed the architecture specifically for the task: it analyses the MRI in three dimensions to find the subtle structural anomalies that characterise an FCD. Autoencoder regularisation improves how well it generalises across scans from different patients and machines.
The largest FCD training dataset of its time
Data scarcity is one of the hardest problems in medical AI. Through extensive collaboration with the Epilepsy Center we compiled what the 2021 paper describes as the largest FCD training dataset to date, covering various types of FCD. Without it the model would only have worked on the cases it had already seen.
More sensitive than the eye
In the prospective validation published in 2021, the model found 7 of 9 FCDs in daily-routine MRIs, where conventional visual analysis found 3. The clinical team assessed it as highly useful for FCD screening in practice.
It keeps learning without forgetting
The system uses a continual learning approach, so it improves as new cases arrive while minimising catastrophic forgetting, the failure mode where training on new data destroys what a model already knew.
What changed
The results
Before
FCD detection depended on a tiny number of global specialists. Manual MRI analysis was slow, subjective and unavailable at most hospitals, and many patients went undiagnosed.
After
An automated screening tool with state-of-the-art detection accuracy, higher sensitivity than conventional visual analysis, peer-reviewed and prospectively validated in clinical practice.
The model reached state-of-the-art results in FCD detection and was validated in clinical practice at the Hamburg Epilepsy Center.
The findings were published in the peer-reviewed journal Epilepsy Research in 2021 and 2024, which puts them in front of the wider medical community rather than leaving them inside one hospital.
The screening no longer depends on whether one of a few specialists is available, which is what decides whether a treatable patient gets the right diagnosis.
Peer-reviewed publications
- Development and prospective clinical validation of a convolutional neural network for automated detection and segmentation of focal cortical dysplasiasChanra V, Chudzinska A, Braniewska N, Silski B, Holst B, Sauvigny T, Stodieck S, Pelzl S, House PM (2024). Epilepsy Research, 202, 107357. DOI 10.1016/j.eplepsyres.2024.107357ResearchGate
- Automated detection and segmentation of focal cortical dysplasias (FCDs) with artificial intelligence: Presentation of a novel convolutional neural network and its prospective clinical validationHouse PM, Kopelyan M, Braniewska N, Silski B, Chudzinska A, Holst B, Sauvigny T, Martens T, Stodieck S, Pelzl S (2021). Epilepsy Research, 172, 106594. DOI 10.1016/j.eplepsyres.2021.106594ResearchGate
Questions about this project
Epilepsy lesion detection on MRI. A critical diagnostic step, identifying FCDs on MRI, depended on a tiny pool of global specialists. Manual analysis was time-consuming, subjective and out of reach for most hospitals, and patients were being missed.
Sensitivity in detecting FCD lesions: 90%. Specificity: 70%. Peer-reviewed in Epilepsy Research, 2021 and 2024: 2 papers. The largest FCD training dataset at publication in 2021: Largest.
3D CNN with autoencoder regularisation. Technology used: 3D convolutional neural networks, Autoencoder regularisation, Medical image segmentation, MRI analysis, Continual learning, Deep learning, Python.
Technology used
Talk to the people who built this
theBlue.ai comes back within one business day.
