IMG-85. Development of AI-based leptomeningeal seeding detection model using anomaly detection algorithm

Leptomeningeal seeding (LMS) represents a clinical hurdle in brain metastasis patients, occurring in 1-37% of cases with median survival of 1-4 months. Current diagnosis relies on invasive cerebrospinal fluid analysis with limited sensitivity, while MRI-based diagnosis has unmet needs due to radiologist fatigue, requirement to check longitudinal series, and small, non-specific patterns of lesions. This study developed an AI-based LMS detection model using AutoDDPM (Automatic Denoising Diffusion Probabilistic Models), a novel anomaly detection algorithm that reconstructs normal brain anatomy and identifies pathological regions through image differences. The training dataset comprised open datasets from FastMRI, IXI dataset, and ATLAS: 700 T1-weighted and 1,339 T1 contrast-enhanced normal brain MRI scans, with external validation performed on 136 LMS cases containing 851 slices from SNUH. The AutoDDPM algorithm employs a two-stage diffusion process: first generating coarse anomaly masks using diffusion models, then reconstructing masked regions and detecting anomalies through differences between synthetic and original images. Three model variants were developed: T1-weighted fine-tuned, T1-contrast supratentorial, and T1-contrast mid-axial, optimized for different slice positions and contrast conditions. Quantitative evaluation using Mean Absolute Error (MAE), Learned Perceptual Image Patch Similarity (LPIPS), and Structural Similarity Index (SSIM) demonstrated acceptable performance across all models. The T1-contrast mid-axial model achieved optimal results with MAE of 0.0193±0.0134, LPIPS of 0.0524±0.0353, and SSIM of 0.869±0.0726. Qualitative analysis revealed successful detection of leptomeningeal enhancement patterns, with the algorithm effectively identifying abnormalities challenging for conventional segmentation approaches. This represents the application of diffusion-based anomaly detection for LMS identification in brain MRI, offering potential for a non-invasive alternative to current diagnostic methods with promising clinical applicability for improving early detection and patient outcomes in brain metastasis management.