Building AI that Learns and Adapts: A Case Study from MRI Diagnostics
Epilepsy affects nearly fifty million people worldwide and remains one of the most common neurological conditions [1]. In many cases, seizures are caused by focal cortical dysplasias (FCDs), subtle abnormalities in brain development that are difficult to detect on standard MRI scans. Their appearance varies widely across patients and often blends into surrounding tissue, which makes reliable identification challenging even for experienced specialists [1,2]. While surgical treatment can be highly effective once these lesions are identified, accurate detection requires a level of expertise that is available only in a limited number of centers.
This diagnostic challenge illustrates a broader technical problem that extends beyond medical imaging. Real-world data is rarely static: conditions change, patterns shift and access to specialized expertise can be limited. AI systems that perform well in controlled settings often struggle once they are deployed in environments where data distributions evolve over time.
In this article, we describe how an AI system was developed to support the detection of subtle abnormalities in MRI data, why continual learning became essential for improving its reliability and what this approach reveals about building AI systems that remain robust as conditions change. While the use case originates in medical imaging, the underlying principles apply to any organization working with complex, variable or expertise-intensive data environments.
Understanding FCDs and Why They Matter
FCDs disrupt normal brain development. Their boundaries may appear blurred, the cortex may thicken, surface folding may deviate from usual patterns and grey-white matter transitions may display subtle intensity changes. Even the well-known transmantle sign (a vertical signal alteration extending from the cortex toward deeper brain structures and considered a classic feature of certain FCD types) does not appear in all patients, which further increases diagnostic uncertainty. These characteristics differ across patients, which means no standard visual template exists.
When specialists detect such lesions, patients often become seizure-free after surgery. When the findings remain hidden, treatment options shrink. This makes early identification essential, although the complexity of the task limits what clinicians can achieve without support.
Transforming MRI Data into Learnable Signals
Together with Evangelisches Krankenhaus Alsterdorf and Dr. Patrick House, we explored how AI could support specialists in early identification of FCDs. The aim was not to automate diagnosis. The emphasis lay in building a system capable of highlighting potentially relevant regions, which would allow experts to focus on the right details faster.
MRI data presents a unique challenge. A single study contains hundreds of slices shaped by different acquisition sequences. T1 sequences reveal anatomical structure with sharp clarity. T2 sequences expose fluid patterns, while FLAIR sequences suppress fluid signals altogether, making abnormalities stand out more clearly.
Before training, the data undergoes segmentation to remove irrelevant structures. Morphometric maps derived using the Huppertz method describe tissue junctions and extensions, which help expose features often invisible in raw scans [3]. The final model input includes T1, FLAIR and two morphometric channels.
The Neural Architecture Behind the Model
The model follows an encoder-decoder design with variational regularization. The encoder compresses the image into an abstract feature space, while the decoder reconstructs target regions by gradually reintroducing spatial detail. This technique helps identify edges, shapes and irregular textures that specialists typically search for manually.
The prospective evaluation showed strong performance. The model identified around 78 percent of the actual FCDs, clearly surpassing what experienced neuroradiologists achieved through visual inspection alone. Specificity remained low, but most false positives were caused by image noise or pathologies unrelated to FCDs. These results highlighted that differences in the data itself play a significant role – an insight that became essential for the next phase of development.
Why Continual Learning Becomes Essential
Static models lose performance as soon as their data environment changes. This is exactly what happens in nearly all real-world applications: scanners are replaced, patient populations differ, acquisition parameters vary. When a model cannot adapt to these changes, its accuracy declines and it requires frequent, costly retraining. Continual learning addresses this fundamental problem by enabling AI systems to grow with new conditions without forgetting what they previously learned.
Since models in real-world applications are often confronted with changing data and conditions, continual learning is becoming increasingly important. It provides a way to build AI systems that remain stable and reliable over time instead of failing whenever the environment shifts.
Continual learning enables a model to incorporate new experiences without overwriting existing knowledge. Biological brains perform this naturally: synapses strengthen when experiences are important and stabilize as memories form. Methods such as Elastic Weight Consolidation (EWC) draw from this principle by protecting parameters that were crucial for earlier tasks.
Replay-based approaches stabilize the model’s “memory” by reintroducing earlier samples, similar to how the human brain reinforces learning through dreaming. Architecture-based methods expand the network to accommodate new tasks while preserving previous capabilities, reflecting principles of neurogenesis.
Applying Continual Learning to FCD Detection
The dataset was reshaped into three sequential distributions, representing real-world variation. Each set contained a small number of FCD MRIs and a much larger number of normal or non-FCD pathological cases.
Two strategies emerged. One used classical training across all aggregated data. The other applied continual learning to new distributions while retaining earlier knowledge. When tested on subsequent datasets, the continual learning approach achieved similar sensitivity with significantly higher specificity. False positives decreased, particularly in frontal and central brain areas. Specificity reached 90 percent when only normal MRIs were considered.
The results show that adaptive models not only work on a technical level but also develop robust and reliable performance under real-world conditions. This genuine adaptive intelligence enables the system to remain stable as data distributions shift, creating a dependable foundation for pre-screening in clinical workflows.
The same principle applies beyond medical imaging: organizations benefit from AI systems that continue to learn, stay reliable and evolve as their environments change.
What Businesses Can Learn from This Use Case
Although this case study grew from a clinical problem, the engineering insights apply to nearly every industry. Logistics environments shift. Market signals fluctuate. Operational sensor data changes with seasons or equipment wear. Static algorithms struggle under such conditions. AI built for continuous adaptation becomes a strategic differentiator.
At theBlue.ai, we design systems that evolve with your data reality. Every project demands its own strategy, architecture and integration pattern. If your organization is exploring AI for processes, our team can help you evaluate the opportunities and design a solution thats perfect for your environment. Together, we can assess your data landscape, identify feasible use cases and design a solution tailored to your processes, infrastructure and long-term strategy.
If you would like to discuss opportunities or explore what is possible, you can reach us anytime through our contact form.
References
[1] House PM, Kopelyan M, Braniewska N, Silski B, Chudzinska A, Holst B, Sauvigny T, Martens T, Stodieck S, Pelzl S. Automated detection and segmentation of focal cortical dysplasias (FCDs) with artificial intelligence: Presentation of a novel convolutional neural network and its prospective clinical validation. Epilepsy Res. 2021 May;172:106594. doi: 10.1016/j.eplepsyres.2021.106594. Epub 2021 Feb 25. PMID: 33677163.
https://pubmed.ncbi.nlm.nih.gov/33677163/
[2] Chanra V, Chudzinska A, Braniewska N, Silski B, Holst B, Sauvigny T, Stodieck S, Pelzl S, House PM. Development and prospective clinical validation of a convolutional neural network for automated detection and segmentation of focal cortical dysplasias. Epilepsy Res. 2024 May;202:107357. doi: 10.1016/j.eplepsyres.2024.107357. Epub 2024 Apr 3. PMID: 38582073.
https://pubmed.ncbi.nlm.nih.gov/38582073/
[3] Huppertz HJ, Grimm C, Fauser S, Kassubek J, Mader I, Hochmuth A, Spreer J, Schulze-Bonhage A. Enhanced visualization of blurred gray-white matter junctions in focal cortical dysplasia by voxel-based 3D MRI analysis. Epilepsy Res. 2005 Oct-Nov;67(1-2):35-50. doi: 10.1016/j.eplepsyres.2005.07.009. Epub 2005 Sep 19. PMID: 16171974. https://pubmed.ncbi.nlm.nih.gov/38582073/
