Deep Learning and Pattern Recognition for Anomaly Detection in Multimodal Medical Imaging With Sensor‐Based Data Fusion

Medical anomaly detection remains a critical challenge due to the heterogeneity of multimodal imaging data and the dynamic nature of physiological sensor readings. Existing fusion‐based approaches, which often rely on shallow concatenation or modality‐specific networks, struggle with cross‐modality inconsistencies, loss of diagnostic cues during integration, and elevated false‐alarm rates, thereby limiting reliable clinical deployment. Moreover, most conventional frameworks fail to jointly capture the spatial complexity of imaging modalities and the temporal dependencies of biosensor data, leading to suboptimal anomaly recognition and classification. To address these limitations, we propose the deep pattern recognition fusion network (DPRFN), a novel deep learning and pattern recognition framework that jointly exploits multimodal medical imaging and physiological sensors through an attention‐driven fusion mechanism for advanced anomaly detection. The architecture integrates ResNet‐50‐based hierarchical feature extraction for spatial representation with a bidirectional gated recurrent unit (Bi‐GRU) to model sequential dependencies in sensor signals. The attention‐guided fusion module ensures optimal cross‐modality harmonization, retaining clinically relevant features while improving sensitivity to subtle anomalies. The framework was validated on the Brain Tumor Segmentation Challenge 2021 (BraTS 2021) dataset (2000+ multimodal magnetic resonance imaging MRI scans with ground truth segmentations) and Medical Information Mart for Intensive Care‐III (MIMIC‐III)/PhysioNet physiological records, which provide rich spatial, volumetric, temporal, and biosignal annotations—enabling robust benchmarking of fusion‐based anomaly detection systems. Experimental results demonstrate significant improvements, achieving accuracy = 97.6%, F1‐score = 96.9%, area under the curve (AUC) = 0.984, and a false‐positive rate = 2.1%, outperforming state‐of‐the‐art baselines by 3%–5% across key metrics. In addition to standard evaluations, we report advanced anomaly detection parameters: structural similarity index (SSIM) = 0.912 for imaging integrity, Matthews correlation coefficient (MCC) = 0.937 for anomaly classification stability, and a diagnostic confidence index (DCI = 0.923), a novel metric introduced to assess the reliability of anomaly detection in clinical decision‐making. Importantly, the reduction in false positives directly translates to fewer unnecessary clinical interventions, highlighting practical medical relevance. The results show that DPRFN delivers a robust and scalable solution for multimodal anomaly detection, significantly reducing diagnostic uncertainty while preserving critical cross‐modality information. Beyond brain imaging, the framework demonstrates high adaptability and reusability, with potential applications in cardiac arrhythmia detection, neurodegenerative disorder monitoring, and other sensor‐driven anomaly detection scenarios. These findings position DPRFN as a next‐generation computational intelligence framework for precision diagnostic imaging and anomaly recognition.