Detecting small, subtle lesions in high-resolution 3D brain MRI remains a challenging and clinically important problem, particularly for identifying epileptogenic abnormalities that may be difficult to visualize with standard reading. The authors position their work within this unmet need and aim to improve automated localization of lesions that are small or otherwise inconspicuous on routine imaging.
The proposed system, called SpyCAT (Spatially Context-Aware Transformer), is a semi-supervised, modeling-based anomaly detection approach. Its design explicitly encodes assumptions about what should be considered normal tissue and what constitutes a relevant deviation from normality; these assumptions form the inductive bias that motivates the architecture and the training strategy. The authors emphasize that SpyCAT is intended to be interpretable in the sense that its anomaly decisions follow from the modeled notion of normalcy rather than from purely discriminative labels alone.
The evaluation uses point-annotations of two subtypes of epileptogenic lesions. The manuscript states that all relevant ethical guidelines were followed, with approval by the Ethics Committee of the medical faculty of Ruhr-University Bochum (approval number 21-7336). The authors confirm that necessary patient consent and institutional approvals were obtained and that identifiers were managed to prevent re-identification by parties outside the research group. The posted preprint provides a DOI and posting date (September 15, 2026). The source indicates that data produced in the study are available upon reasonable request to the authors.
SpyCAT operates on a patch-based representation of 3D MRI volumes. Individual 3D patches are encoded using a vector quantizing variational autoencoder (VQ-VAE) that produces discrete tokens representing patch content. These discrete tokens are then processed by a transformer architecture that is spatially context-aware, enabling the model to reason about relationships among neighboring patches across the volume.
As part of its modeling-based anomaly detection approach, SpyCAT generates counterfactual patches via full 3D convolutions for each patch. This counterfactual generation is integrated into the system to help distinguish expected normal patterns from anomalous deviations. The pipeline therefore combines representation learning (VQ-VAE), sequence modeling with attention (transformer), and 3D convolutional reconstruction to support anomaly localization.
The authors evaluated SpyCAT against reference methods using validation measures that draw on the Metrics Reloaded framework. Evaluation focused on the system's ability to identify and localize the two epileptogenic lesion subtypes represented by point annotations. The manuscript reports qualitative and comparative assessments indicating reliable localization performance and superior performance relative to state-of-the-art reference methods. The source text does not provide numeric performance values, specific metrics, or statistical test results in the abstracted material provided.
According to the report, SpyCAT can reliably identify and localize the lesion types under consideration and outperforms state-of-the-art reference methods on the chosen evaluation tasks. The authors attribute these results to the combination of a patch-based approach, discrete patch tokens from a VQ-VAE, spatially context-aware transformer modeling, and counterfactual patch reconstruction with full 3D convolutions. The preprint emphasizes the value of encoding explicit assumptions about normality into the inductive bias of the anomaly detector.
The preprint states that all data produced in the study are available on reasonable request to the authors. Funding sources reported include the German Federal Ministry of Education and Research (Ministerium für Bildung und Forschung, FKZ 02L19C203) and the Medical Faculty of the Ruhr University Bochum (F1089N-2004). The authors declare no competing interests. The manuscript confirms adherence to reporting guidelines and ethical approvals as noted above.
The work presents a proof-of-concept semi-supervised pipeline tailored to subtle lesion detection in 3D brain MRI, combining generative encoding, discrete-token transformers, and 3D reconstruction. While the authors report that SpyCAT outperforms existing reference methods on their tasks, the source material provided does not include detailed numeric results, external validation cohorts, or implementation release details beyond the statement that data are available on request. Readers seeking reproducibility or detailed performance numbers should consult the full preprint or contact the authors for data and code access as indicated by the preprint entry.
This preprint was posted on medRxiv on September 15, 2026, with DOI: https://doi.org/10.64898/2026.09.14.26363006. The contributing institutions include Ruhr-University Bochum and affiliated clinical departments. The authors report that all necessary ethical approvals and participant consents were obtained and archived.