Explainable detection of goldenhar syndrome using an OD-Mamba backbone for rare craniofacial disorder diagnosis

Goldenhar Syndrome (GS) is a rare congenital craniofacial disorder characterized by asymmetrical facial deformities. It is often underdiagnosed due to the limited availability of clinical data and the absence of automated screening tools. This study establishes a strong foundation for integrating explainable deep learning (DL) models into computerized systems for screening craniofacial anomalies. We employed a publicly available image dataset for the GS detection dataset, comprising 3,550 original and 5,520 augmented craniofacial images, categorized into seven anomaly types. To address the challenges of fine-grained feature localization in rare disorders, we propose YOLO-MvX, a lightweight and adaptive detection model. This model incorporates an OD-Mamba backbone and a Vision Clue Merge (VCM) module to enhance feature fusion and anomaly discrimination. Experiments were conducted with three train/test splits (75/25, 80/20, and 85/15) on both imbalanced and balanced datasets. YOLO-MvX achieved a 96.4% mAP@50, with 94.3% precision and 93.5% recall, outperforming YOLOv9-YOLOv11 by up to 4.3%. Paired t-tests confirmed these improvements as statistically significant [Formula: see text] with large effect sizes. The model showed consistent performance across splits, indicating minimal sensitivity to partitioning, and improved class-wise metrics after balancing, where mean recall increased from 89.5% to 93.3%; mean mAP@50 increased from 92.1% to 95.7%. GradCAM++ provided the most reliable visual explanations (IoU-Heat = 0.87) with enhanced fidelity after balancing. YOLO-MvX also demonstrated efficiency with 4.30M parameters, 10.3 GFLOPs, fast inference speed, and low memory usage. These findings position YOLO-MvX as an efficient and interpretable detector for GS anomaly screening, although external validation under diverse conditions is recommended.