nnU-FCD: A Deep Learning-Based Automatic Segmentation System for Focal Cortical Dysplasia

nnU-FCD is a medical imaging AI system for epilepsy surgery. It uses multi-modal MRI and deep learning architectures (e.g., nnU-Net) to achieve automatic segmentation of Focal Cortical Dysplasia (FCD) lesions. Its core goal is to address the challenges of high difficulty and strong subjectivity in manual FCD identification, improve the accuracy and consistency of preoperative assessment, and ultimately enhance surgical outcomes. The system is released as open-source to promote domain collaboration and clinical application.

Focal Cortical Dysplasia (FCD) is one of the main causes of drug-resistant epilepsy, especially common in pediatric patients. However, FCD lesions are subtle on MRI (e.g., increased cortical thickness, blurred gray-white matter boundaries), leading to multiple challenges in manual diagnosis: lesion features overlapping with normal variations (prone to missed or misdiagnosis), time-consuming interpretation of thin-slice MRI, and limited inter-doctor consistency. Therefore, developing an automated FCD segmentation tool has significant clinical value.

nnU-FCD uses multi-modal MRI inputs (T1-weighted images, T2-weighted images, FLAIR sequences, T1 contrast-enhanced sequences) to fuse tissue characteristic information from different sequences. The system is built based on the nnU-Net framework, whose adaptive features can automatically configure network topology and training strategies according to the dataset without manual parameter tuning. Additionally, combinations of U-Net variants, attention mechanism models, and Transformer architectures are explored to enhance segmentation robustness and scenario flexibility.

The preprocessing pipeline of nnU-FCD includes: image registration (unifying spatial coordinate systems), intensity normalization (eliminating device parameter differences), skull stripping and brain tissue segmentation (focusing on regions of interest), and resampling (unifying voxel spacing). Data augmentation combines spatial transformations (rotation, scaling, elastic deformation) with intensity transformations (noise injection, gamma correction), and adopts special sampling strategies for small lesions to ensure the diversity and anatomical rationality of training samples.

To enhance cross-center generalization capability, nnU-FCD uses multi-center data for training and evaluates real generalization performance through site-level cross-validation (validation sets from hospitals not involved in training). Domain adaptation techniques are also explored to improve adaptation to new data. Evaluation metrics include not only pixel-level metrics such as Dice coefficient and IoU but also lesion-level sensitivity and false positive rate, which align with clinical decision-making needs.

In clinical application, nnU-FCD supports DICOM import, automatic preprocessing, 3D visualization of segmentation results, provides confidence scores to assist doctors' judgment, and can be integrated with neurosurgical navigation systems to aid surgical planning. The open-source release allows free non-commercial use and modification, lowering the entry barrier for institutions with limited medical resources and promoting domain collaboration and system iteration.

Currently, nnU-FCD has limitations: it depends on image quality (artifacts and low resolution affect performance), and the sensitivity to small or complex lesions needs improvement. In the future, we will integrate more modalities such as PET and MEG, develop models dedicated to FCD subtypes, explore unsupervised/semi-supervised learning to reduce annotation dependency, and deeply integrate with electronic medical record systems to achieve full-process assistance.