Peer-reviewed article
Deep learning-based detection of focal cortical dysplasia in children: External validation of the MELD Graph and 3D-nnUNet pipelines.
Andrea Dell'Orco, Enrico De Vita, Felice D'Arco, Annalena Lange, Theodor Rüber et al. ·
Summary
Written by AI, not yet AICS's view · claude-haiku-4-5-20251001This study validates two artificial intelligence models (MELD Graph and 3D-nnUNet) for detecting focal cortical dysplasia, a brain abnormality that causes epilepsy in children. The models were tested on brain MRI scans from 70 children. Both models showed moderate ability to find lesions but missed a significant proportion of cases, suggesting they could help doctors but should not replace expert assessment.
Key points
- MELD Graph achieved 71% sensitivity at patient level; 3D-nnUNet achieved 53% sensitivity
- 3D-nnUNet had higher specificity (86%) with fewer false-positive detections compared to MELD Graph (47%)
- Both models failed to detect a clinically relevant proportion of focal cortical dysplasias
- MELD Graph performance was strongly influenced by accuracy of cortical surface reconstruction
- Models should complement rather than replace expert neuroradiological interpretation
Limitations
This is a single-center retrospective study of 70 children, which limits generalisability. The study evaluates detection accuracy of AI tools rather than clinical outcomes for children with epilepsy, so it cannot tell us whether using these tools actually improves children's health or treatment outcomes.
Abstract
OBJECTIVE: To externally validate two recently developed deep-learning approaches for FCD detection, MELD Graph and 3D-nnUNet, in a pediatric cohort. BACKGROUND: Focal cortical dysplasias (FCDs) are among the most common structural causes of drug-resistant epilepsy in children but are frequently subtle and difficult to detect on MRI. Several automated lesion detection methods have been proposed to support neuroradiological assessment. We externally validated two recently developed deep-learning approaches for FCD detection, MELD Graph and 3D-nnUNet, in a pediatric cohort. METHODS: In this retrospective single-center study, we analyzed brain MRI scans of 70 children evaluated for epilepsy, including 34 MRI-positive patients with suspected FCD and 36 MRI-negative patients, based on primary radiology reports. Both models were applied to 3D T1-weighted and 3D FLAIR images. The detected lesions were reviewed by an experienced pediatric neuroradiologist and classified as true positive, false positive, or false negative. False-positive cases were additionally reviewed using electroclinical findings. RESULTS: MELD Graph and 3D-nnUNet achieved comparable lesion precision (0.86 vs. 0.90) but only moderate recall (0.63 vs. 0.47). At the patient level, MELD Graph demonstrated higher sensitivity (0.71 vs. 0.53), whereas 3D-nnUNet achieved substantially higher specificity (0.86 vs. 0.47) owing to fewer false-positive detections. 3D-nnUNet showed higher sensitivity with the improved FLAIR sequence, whereas MELD Graph performance was strongly influenced by the accuracy of cortical surface reconstruction. Both models failed to detect a clinically relevant proportion of FCDs. CONCLUSION: MELD Graph and 3D-nnUNet generalize to pediatric MRI and represent valuable decision-support tools for highlighting suspicious cortical regions. However, their moderate sensitivity and susceptibility to false-positive findings indicate that they should complement, rather than replace, expert neuroradiological interpretation.